diff --git a/sdk/nexent/core/model_service/__init__.py b/sdk/nexent/core/model_service/__init__.py
new file mode 100644
index 0000000000..8bbf7e4f38
--- /dev/null
+++ b/sdk/nexent/core/model_service/__init__.py
@@ -0,0 +1 @@
+"""Model service contracts and provider adapters."""
diff --git a/sdk/nexent/core/model_service/profiles/LICENSE b/sdk/nexent/core/model_service/profiles/LICENSE
new file mode 100644
index 0000000000..9ef000844a
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 models.dev
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/sdk/nexent/core/model_service/profiles/README.md b/sdk/nexent/core/model_service/profiles/README.md
new file mode 100644
index 0000000000..4797c583c6
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/README.md
@@ -0,0 +1,39 @@
+# Nexent model profile source catalog
+
+## Source and scope
+
+Selected authored TOML files are copied byte-for-byte from
+[models.dev](https://github.com/anomalyco/models.dev), commit
+`e69e1844210e5cfe7c7e7f68d57328bc7451a51c`.
+The original MIT copyright and permission notice is retained in [LICENSE](LICENSE).
+
+The initial inventory contains 21 model files, 21 provider-model files and three
+provider files for Alibaba China, Volcengine and Volcengine Coding Plan. It covers
+selected recent Qwen, GLM, DeepSeek and MiniMax models. This is source material,
+not a verified support list. Pricing remains reference data; this change does
+not implement cost estimation.
+
+## File ownership
+
+- `models/` and `providers/` retain upstream paths, spelling, comments and bytes.
+- `manifest.toml` pins source kinds, paths, hashes and the upstream revision.
+- Nexent normalization and API binding rules belong under `rules/` separately;
+  do not silently rewrite upstream files.
+- Source references are provenance. Saved outlets retain their resolved
+  configuration and do not recompose it when the catalog changes.
+
+## Updating the selection
+
+1. Select one upstream commit and inspect its authored schema and generator.
+2. Copy reviewed files and their complete `base_model` dependencies.
+3. Preserve original bytes and paths; update manifest hashes and revision.
+4. Review every field against the versioned Nexent mapping rules, including
+   fields retained as references or explicitly ignored.
+5. Run integrity, parser, normalization, composition and installed-package
+   checks. Run the reusable real-model matrix before declaring support.
+6. Keep deleted upstream entries locally unless deliberately removed in a Nexent
+   release; do not modify saved outlets automatically.
+
+No upstream TypeScript project, online synchronization or npm dependency is
+required at runtime. Model weight licenses and service terms remain separate
+from the license of these descriptive files.
diff --git a/sdk/nexent/core/model_service/profiles/__init__.py b/sdk/nexent/core/model_service/profiles/__init__.py
new file mode 100644
index 0000000000..b32bda4f07
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/__init__.py
@@ -0,0 +1 @@
+"""Local model profile definitions, loading and resolution."""
diff --git a/sdk/nexent/core/model_service/profiles/_models_dev/__init__.py b/sdk/nexent/core/model_service/profiles/_models_dev/__init__.py
new file mode 100644
index 0000000000..363963c931
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/_models_dev/__init__.py
@@ -0,0 +1 @@
+"""Private parsing of the pinned models.dev authored source format."""
diff --git a/sdk/nexent/core/model_service/profiles/_models_dev/reader.py b/sdk/nexent/core/model_service/profiles/_models_dev/reader.py
new file mode 100644
index 0000000000..0ab4bfb84e
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/_models_dev/reader.py
@@ -0,0 +1,134 @@
+"""Read the complete declared source inventory before semantic normalization."""
+
+import hashlib
+import tomllib
+from dataclasses import dataclass
+from pathlib import Path
+from types import MappingProxyType
+from typing import Literal, Mapping
+
+from pydantic import Field, ValidationError
+
+from nexent.core.model_service.profiles.contracts import AtomicProfileReference, NonBlankString, ProfileContract
+
+
+SourceKind = Literal["model", "provider", "provider_model"]
+
+
+class SourceEntry(ProfileContract):
+    """One source file whose bytes are pinned by the catalog manifest."""
+
+    kind: SourceKind
+    path: NonBlankString
+    content_sha256: NonBlankString
+
+
+class SourceManifest(ProfileContract):
+    """Versioned source inventory; mapping and runtime API bindings are separate."""
+
+    schema_version: int = Field(ge=1, le=1)
+    source_schema_ref: Literal["models_dev_authored_v1"]
+    upstream_url: Literal["https://github.com/anomalyco/models.dev"]
+    upstream_revision: NonBlankString
+    sources: list[SourceEntry] = Field(min_length=1)
+
+
+class SourceReadError(ValueError):
+    """Sanitized source diagnostic without embedding file contents or OS errors."""
+
+    def __init__(self, category: str, file_path: str, source_kind: str = "manifest"):
+        self.category = category
+        self.file_path = file_path
+        self.source_kind = source_kind
+        self.source_path = ""
+        self.rule_id = "source_inventory_v1"
+        super().__init__(f"{category}: {source_kind}: {file_path}")
+
+
+@dataclass(frozen=True)
+class SourceDocument:
+    """Deeply read-only authored document, not a resolved runtime profile."""
+
+    kind: SourceKind
+    key: str
+    file_path: str
+    content_sha256: str
+    data: Mapping[str, object]
+
+
+def _freeze(value: object) -> object:
+    if isinstance(value, dict):
+        return MappingProxyType({key: _freeze(item) for key, item in value.items()})
+    if isinstance(value, list):
+        return tuple(_freeze(item) for item in value)
+    return value
+
+
+def _read(root: Path, path: str, kind: str) -> bytes:
+    try:
+        candidate = (root / path).resolve()
+        if not candidate.is_relative_to(root):
+            raise SourceReadError("SOURCE_PATH_OUTSIDE_ROOT", path, kind)
+        return candidate.read_bytes()
+    except (OSError, RuntimeError):
+        raise SourceReadError("SOURCE_FILE_UNAVAILABLE", path, kind) from None
+
+
+def _parse(data: bytes, path: str, kind: str) -> dict:
+    try:
+        return tomllib.loads(data.decode("utf-8"))
+    except (UnicodeDecodeError, tomllib.TOMLDecodeError):
+        raise SourceReadError("SOURCE_TOML_INVALID", path, kind) from None
+
+
+def _key(entry: SourceEntry) -> str:
+    parts = entry.path.split("/")
+    if entry.kind == "model" and len(parts) >= 3 and parts[0] == "models":
+        return entry.path[len("models/") : -len(".toml")]
+    if entry.kind == "provider" and len(parts) == 3 and parts[0] == "providers" and parts[2] == "provider.toml":
+        return parts[1]
+    if entry.kind == "provider_model" and len(parts) >= 4 and parts[0] == "providers" and parts[2] == "models":
+        return parts[1] + "/" + "/".join(parts[3:])[: -len(".toml")]
+    raise SourceReadError("SOURCE_KIND_PATH_MISMATCH", entry.path, entry.kind)
+
+
+def read_sources(root: Path) -> tuple[SourceDocument, ...]:
+    """Fail the whole inventory on errors; do not expose a partial catalog."""
+    root = root.resolve()
+    raw_manifest = _parse(_read(root, "manifest.toml", "manifest"), "manifest.toml", "manifest")
+    try:
+        manifest = SourceManifest.model_validate(raw_manifest)
+        if len(manifest.upstream_revision) != 40 or any(
+            char not in "0123456789abcdef" for char in manifest.upstream_revision
+        ):
+            raise ValueError("Invalid upstream commit")
+    except (ValidationError, ValueError):
+        raise SourceReadError("SOURCE_MANIFEST_INVALID", "manifest.toml") from None
+
+    documents = []
+    identities = set()
+    paths = set()
+    for entry in manifest.sources:
+        try:
+            AtomicProfileReference(
+                key=entry.path,
+                catalog_revision=manifest.upstream_revision,
+                file_path=entry.path,
+                content_sha256=entry.content_sha256,
+            )
+        except ValidationError:
+            raise SourceReadError("SOURCE_REFERENCE_INVALID", entry.path, entry.kind) from None
+        if not entry.path.endswith(".toml") or entry.path.split("/")[-1] == ".toml":
+            raise SourceReadError("SOURCE_KIND_PATH_MISMATCH", entry.path, entry.kind)
+        key = _key(entry)
+        identity = (entry.kind, key)
+        if entry.path in paths or identity in identities:
+            raise SourceReadError("SOURCE_DUPLICATE", entry.path, entry.kind)
+        paths.add(entry.path)
+        identities.add(identity)
+        data = _read(root, entry.path, entry.kind)
+        if hashlib.sha256(data).hexdigest() != entry.content_sha256:
+            raise SourceReadError("SOURCE_DIGEST_MISMATCH", entry.path, entry.kind)
+        parsed = _parse(data, entry.path, entry.kind)
+        documents.append(SourceDocument(entry.kind, key, entry.path, entry.content_sha256, _freeze(parsed)))
+    return tuple(documents)
diff --git a/sdk/nexent/core/model_service/profiles/contracts.py b/sdk/nexent/core/model_service/profiles/contracts.py
new file mode 100644
index 0000000000..1e80413cca
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/contracts.py
@@ -0,0 +1,131 @@
+"""Strict profile primitives, independent of provider clients and persistence."""
+
+import json
+from typing import Annotated, Literal, Self
+
+from pydantic import BaseModel, ConfigDict, Field, StringConstraints, model_validator
+
+
+SupportStatus = Literal["supported", "unsupported", "unknown"]
+NonBlankString = Annotated[str, StringConstraints(pattern=r"\S")]
+PositiveTokens = Annotated[int, Field(gt=0)]
+
+
+class ProfileContract(BaseModel):
+    """Reject undeclared fields and coercion at the profile boundary."""
+
+    model_config = ConfigDict(strict=True, extra="forbid", frozen=True)
+
+
+class ContextCapabilities(ProfileContract):
+    """Effective service limits; missing limits remain unknown."""
+
+    combined_window: PositiveTokens | None = None
+    input_limit: PositiveTokens | None = None
+    output_limit: PositiveTokens | None = None
+
+    @model_validator(mode="after")
+    def validate_limits(self) -> Self:
+        if self.combined_window is not None:
+            for limit in (self.input_limit, self.output_limit):
+                if limit is not None and limit > self.combined_window:
+                    raise ValueError("Service limits cannot exceed the combined context window")
+        return self
+
+
+class ToggleControl(ProfileContract):
+    """Explicitly supported thinking switch values, not the current setting."""
+
+    mode: Literal["toggle"] = "toggle"
+    parameter_key: Literal["enable_thinking"] = "enable_thinking"
+    values: list[bool] = Field(min_length=1)
+
+    @model_validator(mode="after")
+    def validate_values(self) -> Self:
+        if len(set(self.values)) != len(self.values):
+            raise ValueError("Toggle values must be unique")
+        return self
+
+
+class EffortControl(ProfileContract):
+    """Provider-declared effort values; no universal effort enum is inferred."""
+
+    mode: Literal["effort"] = "effort"
+    parameter_key: Literal["reasoning_effort"] = "reasoning_effort"
+    values: list[NonBlankString] = Field(min_length=1)
+
+    @model_validator(mode="after")
+    def validate_values(self) -> Self:
+        if len(set(self.values)) != len(self.values):
+            raise ValueError("Effort values must be unique")
+        return self
+
+
+class BudgetTokensControl(ProfileContract):
+    """Token bounds and explicitly declared sentinel values."""
+
+    mode: Literal["budget_tokens"] = "budget_tokens"
+    parameter_key: Literal["reasoning_budget_tokens"] = "reasoning_budget_tokens"
+    min: Annotated[int, Field(ge=0)]
+    max: Annotated[int, Field(ge=0)]
+    special_values: list[int] = Field(default_factory=list)
+
+    @model_validator(mode="after")
+    def validate_range(self) -> Self:
+        if self.min > self.max:
+            raise ValueError("Budget minimum cannot exceed maximum")
+        if len(set(self.special_values)) != len(self.special_values):
+            raise ValueError("Special budget values must be unique")
+        return self
+
+
+ReasoningControl = Annotated[ToggleControl | EffortControl | BudgetTokensControl, Field(discriminator="mode")]
+
+
+class ReasoningCapabilities(ProfileContract):
+    """Capability declarations with unknown distinct from unsupported."""
+
+    status: SupportStatus = "unknown"
+    controls: list[ReasoningControl] | None = None
+    response_paths: list[NonBlankString] = Field(default_factory=list)
+
+    @model_validator(mode="after")
+    def validate_controls(self) -> Self:
+        if self.status == "unsupported" and (self.controls or self.response_paths):
+            raise ValueError("Unsupported reasoning cannot declare controls or response paths")
+        if len(set(self.response_paths)) != len(self.response_paths):
+            raise ValueError("Reasoning response paths must be unique")
+        if self.controls is not None:
+            modes = [control.mode for control in self.controls]
+            keys = [control.parameter_key for control in self.controls]
+            if len(set(modes)) != len(modes) or len(set(keys)) != len(keys):
+                raise ValueError("Reasoning control modes and parameter keys must be unique")
+            encoded = json.dumps(
+                [control.model_dump(mode="json") for control in self.controls],
+                ensure_ascii=False,
+                separators=(",", ":"),
+            ).encode("utf-8")
+            if len(encoded) > 64 * 1024:
+                raise ValueError("Reasoning controls exceed the 64 KiB UTF-8 limit")
+        return self
+
+
+class AtomicProfileReference(ProfileContract):
+    """Provenance only; runtime loading never recomposes from this reference."""
+
+    key: NonBlankString
+    catalog_revision: NonBlankString
+    file_path: NonBlankString
+    content_sha256: Annotated[str, StringConstraints(pattern=r"^[a-f0-9]{64}$")]
+
+    @model_validator(mode="after")
+    def validate_source_path(self) -> Self:
+        parts = self.file_path.split("/")
+        if (
+            "\\" in self.file_path
+            or ":" in self.file_path
+            or any(ord(character) < 32 for character in self.file_path)
+            or any(part in ("", ".", "..") for part in parts)
+        ):
+            raise ValueError("Source paths must be canonical relative POSIX paths")
+        return self
diff --git a/sdk/nexent/core/model_service/profiles/manifest.toml b/sdk/nexent/core/model_service/profiles/manifest.toml
new file mode 100644
index 0000000000..5ef8ac216d
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/manifest.toml
@@ -0,0 +1,229 @@
+schema_version = 1
+source_schema_ref = "models_dev_authored_v1"
+upstream_url = "https://github.com/anomalyco/models.dev"
+upstream_revision = "e69e1844210e5cfe7c7e7f68d57328bc7451a51c"
+
+[[sources]]
+kind = "model"
+path = "models/alibaba/qwen3.5-27b.toml"
+content_sha256 = "15e1c86123b937d3152f69e87650f3b3bbfd9600f8bbfbc5361a6f5e9f69fc16"
+
+[[sources]]
+kind = "model"
+path = "models/alibaba/qwen3.5-35b-a3b.toml"
+content_sha256 = "50a139d95791c9c5e145f4a0f115f18c3d38840202c7683e8861f05fe4ffc6be"
+
+[[sources]]
+kind = "model"
+path = "models/alibaba/qwen3.5-397b-a17b.toml"
+content_sha256 = "d064d5959f3dca64beb7108a38280877719e89709db7775f932a5ee791e3924e"
+
+[[sources]]
+kind = "model"
+path = "models/alibaba/qwen3.5-flash.toml"
+content_sha256 = "39a4329d9772432e58cf92044f143bd7e1445b434ee65830acad958b72322d5c"
+
+[[sources]]
+kind = "model"
+path = "models/alibaba/qwen3.5-plus.toml"
+content_sha256 = "15c6b3b7f334219ee6d43c76d1a004e57469e4a4e780e4b254c91474d5cb1bd4"
+
+[[sources]]
+kind = "model"
+path = "models/alibaba/qwen3.8-27b.toml"
+content_sha256 = "5a7c7ec1b513c3268756c02fa240f09ba438bce594cbb7d34ebb4f71c4a448e0"
+
+[[sources]]
+kind = "model"
+path = "models/alibaba/qwen3.8-2.4t-a95b.toml"
+content_sha256 = "067a17c3a34382a3e286414e8f6ef4ad31844716ed85af41fdab7061ba44add4"
+
+[[sources]]
+kind = "model"
+path = "models/alibaba/qwen3.8-flash.toml"
+content_sha256 = "7fc097dbfdcbcf827fc28c17e23bafda6d695ce5d16d96fb65612e4cae581dcc"
+
+[[sources]]
+kind = "model"
+path = "models/alibaba/qwen3.8-max.toml"
+content_sha256 = "ee87c9af3b4188b86dc8ee3d045e9a18a0b611b5db1e47cfc3f88b1672164634"
+
+[[sources]]
+kind = "model"
+path = "models/alibaba/qwen3-coder-next.toml"
+content_sha256 = "1f4175a40524b382f1e9efac28e4454b477a2f4ccf5a460134b9ed12cb829229"
+
+[[sources]]
+kind = "model"
+path = "models/zhipuai/glm-5.1.toml"
+content_sha256 = "8c7069a86e0ef764d2a1d0ebb30156abed0cf7b27e5b972c224a5cd9aee27b40"
+
+[[sources]]
+kind = "model"
+path = "models/zhipuai/glm-5.2.toml"
+content_sha256 = "6ed303f1feeaed5496e919b6c7ae999e2743505a22ed97062b5c833adf37ce13"
+
+[[sources]]
+kind = "model"
+path = "models/zhipuai/glm-5.3.toml"
+content_sha256 = "8022f2c087c0ad8bdf9f8ed4fe950fda5f7f37ef2a121c577e87fbe28d159c29"
+
+[[sources]]
+kind = "model"
+path = "models/zhipuai/glm-5.3-flash.toml"
+content_sha256 = "7cdb7a1cfea9ab7e7c3e3392f0fdad0ac9c3f9c92e2f71a2d109788c5d02f843"
+
+[[sources]]
+kind = "model"
+path = "models/deepseek/deepseek-v4-flash.toml"
+content_sha256 = "02ae90fe9be99047c373f3862afe72367c7a96a7aef7f43864c0bcd9e22da194"
+
+[[sources]]
+kind = "model"
+path = "models/deepseek/deepseek-v4-pro.toml"
+content_sha256 = "bd08e3df5404ff5f795ac91b44ef34ac7c29a6911de0bd9fded3739501e47996"
+
+[[sources]]
+kind = "model"
+path = "models/deepseek/deepseek-v4-flash-0731.toml"
+content_sha256 = "cb0133fb9f376da44fc6ee512c8a3ba0399626bb026805f608f265e504511c4e"
+
+[[sources]]
+kind = "model"
+path = "models/deepseek/deepseek-v4-pro-0813.toml"
+content_sha256 = "e005b286813f8bcf39ecc4397b36a99895532c97bbc2c83f21a94dbaa85455ea"
+
+[[sources]]
+kind = "model"
+path = "models/deepseek/deepseek-v4.1-flash.toml"
+content_sha256 = "ef594fbed1c96863ff728e31530dbc1987d8862e36c5f1b5f08e5376d631aee0"
+
+[[sources]]
+kind = "model"
+path = "models/minimax/MiniMax-M2.7.toml"
+content_sha256 = "c63b15b0f281e217f466b973439cfd1876b23d636288cd4f7102ef28a57df761"
+
+[[sources]]
+kind = "model"
+path = "models/minimax/MiniMax-M3.toml"
+content_sha256 = "a1d815dc34328bd0719fc557794d696205a7506a06515f8e2d84c24f3bce1b1e"
+
+[[sources]]
+kind = "provider"
+path = "providers/alibaba-cn/provider.toml"
+content_sha256 = "0060172cf93a8db8b8427021030a3ad5c388c919f7fc47df9723ba8b08d5f54b"
+
+[[sources]]
+kind = "provider"
+path = "providers/volcengine/provider.toml"
+content_sha256 = "b504a2ab81fccb036105dde4a626f486d82e37ceaf348249baaf0b5c8e234377"
+
+[[sources]]
+kind = "provider"
+path = "providers/volcengine-coding-plan/provider.toml"
+content_sha256 = "15439498a42304ac424f73810fdf8593ee87d19c05bff271c64383afe6b06ee8"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/alibaba-cn/models/qwen3.5-397b-a17b.toml"
+content_sha256 = "59342059889eac4ab4829384b954db58c72104f9ffc682441d7b39b319b80fb9"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/alibaba-cn/models/qwen3.5-flash.toml"
+content_sha256 = "48d81041f3babeb002b73184e151df0bb4304b1679137c12e396951bbe9c16fe"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/alibaba-cn/models/qwen3.5-plus.toml"
+content_sha256 = "5bd06ed609ab5ceeeecf8012a4ea587d5365fcd49c17f10bfb729f1a436fee0e"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/alibaba-cn/models/qwen3.8-flash.toml"
+content_sha256 = "f496c5f293e2196f69bcd2f705cbb12548f14a1923d6b35d10400cbc7a1c817b"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/alibaba-cn/models/qwen3.8-max.toml"
+content_sha256 = "1b1b47992952fb615bdce07fdeea93bdf445b36375acae6e5e812e6aecb9bd08"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/alibaba-cn/models/glm-5.1.toml"
+content_sha256 = "f49b1f19c24db0d339b7bafb8c909c15bdaf5dc3765f9befef66b1166e10c520"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/alibaba-cn/models/glm-5.2.toml"
+content_sha256 = "dbfbc406418c60aa33f33eed147f8d47a98f544b6f5cb50bf96e20fb5c38a2c2"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/alibaba-cn/models/glm-5.3.toml"
+content_sha256 = "6ce1f84a6cd1af13f823e846b750e4f39d90aa4f008178091a118f88765dfe2f"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/alibaba-cn/models/deepseek-v4-flash.toml"
+content_sha256 = "b7f6b5d17332d596883c8d5ac7de1c9e885dc4df8e9fc2ea996420e202a7febb"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/alibaba-cn/models/deepseek-v4-pro.toml"
+content_sha256 = "db39defdd8ed98cf27ccb03f9884431c1e52f1b5d6a88591888d5801bb1e0ac3"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/alibaba-cn/models/deepseek-v4.1-flash.toml"
+content_sha256 = "fe4c026f77d2a2b837d390dcef3da521b5dc3298eafb1e06199360cc0c6cd5d0"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/alibaba-cn/models/MiniMax/MiniMax-M2.7.toml"
+content_sha256 = "d14f335b07a27d36e697d350d2aac28d6f9215e6ad4c2912ff9fe8bd5ac7c2f4"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/volcengine/models/deepseek-v4-flash-ga-260731.toml"
+content_sha256 = "8012bc8a96ff1a96a56b64721e5e144d4ff997b70985c34483b95c7469de13c1"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/volcengine/models/deepseek-v4-pro-ga-260813.toml"
+content_sha256 = "444e6d130d11e3132d34e3afe4d878f96120beef51db5946f206d96bc8dfdec8"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/volcengine/models/glm-5-2-260617.toml"
+content_sha256 = "092fcf694c283b90a71cde982d4231d63e4a53544665d4a481e306b4cb1bed1a"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/volcengine/models/glm-5-3-flash-260828.toml"
+content_sha256 = "34b35e1742e663e54d1231baa5282fc1497b255d60764df21e018c8a78da04a2"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/volcengine-coding-plan/models/deepseek-v4-flash.toml"
+content_sha256 = "be24a803ea861218c4ad9fc60025a424cb275b95e7c02507699776ce37869a99"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/volcengine-coding-plan/models/deepseek-v4-pro.toml"
+content_sha256 = "b359870df413e10509083d9b56a62453c530e0f600ca5628ae3279043df2ee57"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/volcengine-coding-plan/models/glm-5.3.toml"
+content_sha256 = "6fb6ebfb16a106cfb22e5f887326d661b96c1b67df4f826fdf6d02cceded4782"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/volcengine-coding-plan/models/glm-5.3-flash.toml"
+content_sha256 = "8ca9644166cb0cdc479049a0f7467a6cebded93cd0d3bef00a66e462b77111f3"
+
+[[sources]]
+kind = "provider_model"
+path = "providers/volcengine-coding-plan/models/minimax-m3.toml"
+content_sha256 = "c2998f64873231e71f62be2da7983afd4c1af76e7a129f232007e65a58c24e24"
diff --git a/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3-coder-next.toml b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3-coder-next.toml
new file mode 100644
index 0000000000..cce2c7bb62
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3-coder-next.toml
@@ -0,0 +1,27 @@
+# https://qwen.ai/blog?id=qwen3-coder-next
+# https://huggingface.co/Qwen/Qwen3-Coder-Next
+# https://www.qwencloud.com/models/qwen3-coder-next
+name = "Qwen3 Coder Next"
+description = "Open-weight Qwen coding model for agents, repository edits, and multi-turn tool use"
+family = "qwen"
+release_date = "2026-02-03"
+last_updated = "2026-02-03"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2025-09"
+open_weights = true
+
+[limit]
+context = 262_144
+output = 65_536
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3-Coder-Next"
diff --git a/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-27b.toml b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-27b.toml
new file mode 100644
index 0000000000..93ad247dd8
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-27b.toml
@@ -0,0 +1,29 @@
+name = "Qwen3.5 27B"
+description = "Qwen vision-language model for visual reasoning, documents, and agent tasks"
+family = "qwen"
+release_date = "2026-02-23"
+last_updated = "2026-02-23"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+
+[limit]
+context = 262_144
+output = 65_536
+
+[modalities]
+input = ["text", "image", "video", "audio"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3.5-27B"
+
+[[benchmarks]]
+name = "SWE-Bench Verified"
+score = 72.4
+metric = "resolved"
+source = "https://huggingface.co/Qwen/Qwen3.5-27B"
diff --git a/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-35b-a3b.toml b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-35b-a3b.toml
new file mode 100644
index 0000000000..36700f959e
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-35b-a3b.toml
@@ -0,0 +1,23 @@
+name = "Qwen3.5 35B-A3B"
+description = "Qwen vision-language model for visual reasoning, documents, and agent tasks"
+family = "qwen"
+release_date = "2026-02-23"
+last_updated = "2026-02-23"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+
+[limit]
+context = 262_144
+output = 65_536
+
+[modalities]
+input = ["text", "image", "video", "audio"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3.5-35B-A3B"
diff --git a/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-397b-a17b.toml b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-397b-a17b.toml
new file mode 100644
index 0000000000..103227c16b
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-397b-a17b.toml
@@ -0,0 +1,29 @@
+name = "Qwen3.5 397B-A17B"
+description = "Large open Qwen multimodal MoE for visual agents and long technical tasks"
+family = "qwen"
+release_date = "2026-02-15"
+last_updated = "2026-02-15"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+
+[limit]
+context = 262_144
+output = 65_536
+
+[modalities]
+input = ["text", "image", "video", "audio"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3.5-397B-A17B"
+
+[[benchmarks]]
+name = "SWE-Bench Verified"
+score = 76.4
+metric = "resolved"
+source = "https://huggingface.co/Qwen/Qwen3.5-397B-A17B"
diff --git a/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-flash.toml b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-flash.toml
new file mode 100644
index 0000000000..800ff85f43
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-flash.toml
@@ -0,0 +1,22 @@
+# https://help.aliyun.com/en/model-studio/qwen3-5-flash
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
+name = "Qwen3.5 Flash"
+description = "Qwen vision-language model for visual reasoning, documents, and agent tasks"
+family = "qwen"
+release_date = "2026-02-23"
+last_updated = "2026-02-23"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+
+[limit]
+context = 1_000_000
+output = 65_536
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]
diff --git a/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-plus.toml b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-plus.toml
new file mode 100644
index 0000000000..7f04f4c2fe
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.5-plus.toml
@@ -0,0 +1,19 @@
+name = "Qwen3.5 Plus"
+description = "Qwen vision-language model for visual reasoning, documents, and agent tasks"
+family = "qwen"
+release_date = "2026-02-16"
+last_updated = "2026-02-16"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+knowledge = "2025-04"
+open_weights = false
+
+[limit]
+context = 1_000_000
+output = 65_536
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]
diff --git a/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.8-2.4t-a95b.toml b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.8-2.4t-a95b.toml
new file mode 100644
index 0000000000..b30a03ea6b
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.8-2.4t-a95b.toml
@@ -0,0 +1,33 @@
+# Sources (accessed 2026-08-16):
+# https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
+# https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/README.md
+# https://qwen.ai/blog?id=qwen3.8
+# https://openrouter.ai/qwen/qwen3.8-2.4t-a95b
+# Open-weight twin of Qwen3.8 Max: text-only, thinking always on,
+# reasoning_effort low|medium|xhigh (default xhigh). Native context 262K,
+# extensible to ~1.01M. Distinct from closed multimodal qwen3.8-max.
+
+name = "Qwen3.8 2.4T A95B"
+description = "Open-weight sparse MoE (2.4T total, 95B active), the open-weight twin of Qwen3.8 Max for coding, research, complex reasoning, and agentic workflows"
+family = "qwen"
+release_date = "2026-08-12"
+last_updated = "2026-08-12"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+license = "qwen3.8-max"
+
+[limit]
+context = 262_144
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
diff --git a/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.8-27b.toml b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.8-27b.toml
new file mode 100644
index 0000000000..4517f55223
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.8-27b.toml
@@ -0,0 +1,236 @@
+# Benchmark caveats and evaluation harnesses: https://huggingface.co/Qwen/Qwen3.8-27B#evaluation
+# Sources (accessed 2026-08-15):
+# https://huggingface.co/Qwen/Qwen3.8-27B
+# https://huggingface.co/api/models/Qwen/Qwen3.8-27B
+# https://qwen.ai/blog?id=qwen3.8
+# Hub lastModified 2026-08-14T15:00:01Z is the open-weight drop.
+# Do not use Hub createdAt 2026-08-05 (staged countdown page).
+
+name = "Qwen3.8 27B"
+description = "Dense 27B vision-language model for coding, agent tasks, and image and video understanding"
+family = "qwen"
+release_date = "2026-08-14"
+last_updated = "2026-08-14"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+
+[limit]
+context = 262_144
+output = 32_768
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "SWE-bench Pro"
+score = 61.7
+metric = "resolved"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 73
+metric = "score"
+version = "2.1"
+harness = "Terminus"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "NL2Repo"
+score = 42.3
+metric = "score"
+harness = "Claude Code"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 42.2
+metric = "score"
+version = "1.1"
+harness = "Claude Code"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "QwenSWEBench"
+score = 79
+metric = "avg@3"
+harness = "Claude Code"
+dataset = "internal"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "CoWorkBench"
+score = 70.7
+metric = "score"
+dataset = "internal"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "JobBench"
+score = 33.4
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "Agents' Last Exam"
+score = 20.4
+metric = "pass@1"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "Agents' Last Exam"
+score = 42.9
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "IFBench"
+score = 79.5
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "GPQA Diamond"
+score = 89.2
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 30.8
+metric = "score"
+variant = "without tools; GPT-4o judge"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "LiveCodeBench"
+score = 90.3
+metric = "score"
+version = "6"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "OSWorld-Verified"
+score = 84.3
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "WebArena-Verified"
+score = 64.8
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+harness = "OSWorld scaffold; official WebArena-Verified grader"
+
+[[benchmarks]]
+name = "AndroidWorld"
+score = 81.9
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "SWE-MM"
+score = 38.6
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "RealWorldQA"
+score = 85.9
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "ERQA"
+score = 65.5
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "RecreationBench"
+score = 47.1
+metric = "score"
+dataset = "internal"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "ClawEval-MM"
+score = 57.4
+metric = "pass@3"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "ClawEval-MM"
+score = 56.9
+metric = "average score"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "Vision2Web"
+score = 62.9
+metric = "score"
+harness = "Claude Code"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+variant = "gpt-5.4-2026-03-05 judge"
+
+[[benchmarks]]
+name = "MathVision"
+score = 90
+metric = "score"
+variant = "without code interpreter"
+dataset = "corrected annotations"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "MathVision"
+score = 94.6
+metric = "score"
+variant = "with code interpreter"
+dataset = "corrected annotations"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "BabyVision"
+score = 65.7
+metric = "score"
+variant = "without code interpreter"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "BabyVision"
+score = 85.6
+metric = "score"
+variant = "with code interpreter"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "CharXiv Reasoning"
+score = 83.7
+metric = "score"
+variant = "without code interpreter"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+dataset = "corrected annotations"
+
+[[benchmarks]]
+name = "CharXiv Reasoning"
+score = 90.2
+metric = "score"
+variant = "with code interpreter"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
+dataset = "corrected annotations"
+
+[[benchmarks]]
+name = "OmniDocBench"
+score = 91.1
+metric = "score"
+version = "1.5"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"
diff --git a/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.8-flash.toml b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.8-flash.toml
new file mode 100644
index 0000000000..2ee8a5ff53
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.8-flash.toml
@@ -0,0 +1,19 @@
+# Source: https://www.qwencloud.com/models/qwen3.8-flash
+name = "Qwen3.8 Flash"
+description = "Qwen vision-language model for visual reasoning, documents, and agent tasks"
+family = "qwen"
+release_date = "2026-08-26"
+last_updated = "2026-08-26"
+attachment = true
+reasoning = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+[limit]
+context = 1_000_000
+output = 131_072
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]
diff --git a/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.8-max.toml b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.8-max.toml
new file mode 100644
index 0000000000..76a956fd3f
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/alibaba/qwen3.8-max.toml
@@ -0,0 +1,266 @@
+# These are the Qwen3.8-Max column of the official model card, not a claim about a separately evaluated local checkpoint.
+# Sources (accessed 2026-08-06):
+# https://www.qwencloud.com/models/qwen3.8-max
+# https://www.qianwenai.com/models/qwen3.8-max
+# https://help.aliyun.com/zh/model-studio/qwen3-8-max
+# https://www.alibabacloud.com/help/en/model-studio/qwen3-8-max
+# https://help.aliyun.com/zh/model-studio/pdf-understanding
+# https://platform.qianwenai.com/docs/developer-guides/tool-calling/pdf-understanding
+# https://docs.qwencloud.com/token-plan/personal/token-plan-personal-overview
+# https://help.aliyun.com/zh/model-studio/token-plan-personal-overview
+# https://help.aliyun.com/en/model-studio/token-plan-personal-overview
+# https://docs.qwencloud.com/developer-guides/getting-started/text-generation-models
+# https://docs.qwencloud.com/developer-guides/text-generation/thinking
+# https://docs.qwencloud.com/developer-guides/clients-and-developer-tools/opencode
+# https://platform.qianwenai.com/docs/developer-guides/clients-and-developer-tools/opencode
+# https://qwen.ai/blog?id=qwen3.8
+# PDF input: Model Studio / 千问AI docs list only qwen3.8-max under PDF理解
+# (type:file / file_url|file_data). Model pages list Image/Text/Video badges
+# and separately list PDF理解 as a Completions built-in tool. Beijing-region
+# availability note on help.aliyun.com; lab capability still includes pdf.
+
+name = "Qwen3.8 Max"
+description = "2.4-trillion-parameter MoE flagship for coding, professional work, multimodal understanding, and long-horizon agentic workflows"
+family = "qwen"
+release_date = "2026-08-03"
+last_updated = "2026-08-03"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = false
+
+[limit]
+context = 1_000_000
+output = 131_072
+
+[modalities]
+input = ["text", "image", "video", "pdf"]
+output = ["text"]
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 86.6
+metric = "avg@10"
+version = "2.1"
+harness = "Claude Code"
+variant = "5h timeout"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 67.7
+metric = "resolved"
+harness = "Claude Code"
+dataset = "Qwen refined and corrected task set"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 56.6
+metric = "score"
+version = "1.1"
+harness = "Claude Code"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "NL2Repo"
+score = 55.9
+metric = "score"
+harness = "Claude Code"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "FrontierSWE"
+score = 73.5
+metric = "dominance score"
+date = "2026-08-03"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "MLS-Bench-Lite"
+score = 41
+metric = "score"
+harness = "Claude Code"
+variant = "5h timeout"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "PaperBench"
+score = 93
+metric = "score"
+harness = "BasicAgent"
+variant = "Code-Dev; 3 runs; 12h timeout; Opus 4.6 judge"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "AndroidBench"
+score = 75.1
+metric = "avg@3"
+dataset = "95 public tasks"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "QwenSWEBench"
+score = 80.7
+metric = "avg@3"
+harness = "Claude Code"
+dataset = "internal"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "QwenQoderBench"
+score = 58.4
+metric = "avg@5"
+harness = "Claude Code"
+dataset = "internal"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "QwenReactBench"
+score = 1724
+metric = "Elo"
+dataset = "internal"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "QwenSVGBench"
+score = 1713
+metric = "Elo"
+dataset = "internal"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "CoWorkBench"
+score = 74.8
+metric = "score"
+dataset = "internal"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "WorkSpaceBench"
+score = 67.7
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "JobBench"
+score = 53.4
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "SkillsBench"
+score = 70.2
+metric = "avg@3"
+version = "1.1"
+dataset = "87 tasks"
+harness = "OpenCode"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "Agents' Last Exam"
+score = 27
+metric = "pass@1"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "Agents' Last Exam"
+score = 52.4
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "AutomationBench"
+score = 27.3
+metric = "pass@1"
+dataset = "600 public tasks"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "Toolathlon-Verified"
+score = 72.5
+metric = "pass@1"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "WideSearch"
+score = 81.9
+metric = "average item F1"
+harness = "Qwen-Agent"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 56.2
+metric = "score"
+variant = "with tools"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 43.6
+metric = "score"
+variant = "without tools"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "GPQA Diamond"
+score = 92.6
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "IFBench"
+score = 82.8
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "$OneMillion-Bench"
+score = 52.5
+metric = "expert score"
+variant = "gemini-3.1-pro-preview judge"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "HealthBench"
+score = 60.2
+metric = "score"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "PLawBench"
+score = 73.2
+metric = "score"
+variant = "gemini-3.1-pro-preview judge"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "PRBench"
+score = 57.6
+metric = "score"
+variant = "legal"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "PRBench"
+score = 58.3
+metric = "score"
+variant = "finance"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "MRCR"
+score = 92.9
+metric = "score"
+version = "2"
+dataset = "256K; 8-needle"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
+
+[[benchmarks]]
+name = "LongBench"
+score = 66.3
+metric = "score"
+version = "2"
+source = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"
diff --git a/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4-flash-0731.toml b/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4-flash-0731.toml
new file mode 100644
index 0000000000..61479ce86a
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4-flash-0731.toml
@@ -0,0 +1,105 @@
+name = "DeepSeek V4 Flash 0731"
+description = "Official DeepSeek V4 Flash release with enhanced agentic capabilities and integrated DSpark speculative decoding"
+family = "deepseek-flash"
+release_date = "2026-07-31"
+last_updated = "2026-07-31"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2025-05"
+open_weights = true
+license = "MIT"
+
+[limit]
+context = 1_000_000
+output = 384_000
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 82.7
+metric = "pass@1"
+variant = "max"
+version = "2.1"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731"
+
+[[benchmarks]]
+name = "NL2Repo"
+score = 54.2
+metric = "resolve rate"
+variant = "max effort"
+harness = "DeepSeek Harness minimal mode"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
+
+[[benchmarks]]
+name = "CyberGym"
+score = 76.7
+metric = "score"
+variant = "max effort"
+harness = "DeepSeek Harness minimal mode"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 54.4
+metric = "resolve rate"
+variant = "max effort"
+harness = "DeepSeek Harness minimal mode"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
+
+[[benchmarks]]
+name = "Toolathlon-Verified"
+score = 70.3
+metric = "score"
+variant = "max effort"
+harness = "DeepSeek Harness minimal mode"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
+
+[[benchmarks]]
+name = "Agents' Last Exam"
+score = 25.2
+metric = "score"
+variant = "max effort"
+harness = "DeepSeek Harness minimal mode"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
+
+[[benchmarks]]
+name = "AutomationBench"
+score = 25.1
+metric = "success rate"
+variant = "max effort"
+dataset = "public"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
+
+[[benchmarks]]
+name = "DSBench-FullStack"
+score = 68.7
+metric = "score"
+variant = "max effort"
+dataset = "internal"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
+
+[[benchmarks]]
+name = "DSBench-Hard"
+score = 59.6
+metric = "score"
+variant = "max effort"
+dataset = "internal"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
diff --git a/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4-flash.toml b/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4-flash.toml
new file mode 100644
index 0000000000..5f521d77d2
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4-flash.toml
@@ -0,0 +1,183 @@
+# Added benchmarks are from the original V4 preview instruct-model comparison, not the later 0731/0813 releases.
+name = "DeepSeek V4 Flash"
+description = "Fast DeepSeek V4 lane for economical reasoning, coding, and long-context work"
+family = "deepseek-flash"
+release_date = "2026-04-24"
+last_updated = "2026-04-24"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2025-05"
+open_weights = true
+
+[limit]
+context = 1_000_000
+output = 384_000
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "SWE-Bench Verified"
+score = 79
+metric = "resolved"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "MMLU-Pro"
+score = 86.2
+metric = "EM"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "SimpleQA-Verified"
+score = 34.1
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "Chinese SimpleQA"
+score = 78.9
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "GPQA Diamond"
+score = 88.1
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 34.8
+metric = "pass@1"
+variant = "preview checkpoint; max effort; without tools"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "LiveCodeBench"
+score = 91.6
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "Codeforces"
+score = 3052
+metric = "rating"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "HMMT"
+score = 94.8
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+dataset = "February 2026"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "IMOAnswerBench"
+score = 88.4
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "MathArena Apex"
+score = 33
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "MathArena Apex Shortlist"
+score = 85.7
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "MRCR"
+score = 78.7
+metric = "MMR"
+variant = "preview checkpoint; max effort"
+dataset = "1M context"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "CorpusQA"
+score = 60.5
+metric = "accuracy"
+variant = "preview checkpoint; max effort"
+dataset = "1M context"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 56.9
+metric = "accuracy"
+variant = "preview checkpoint; max effort"
+version = "2.0"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 52.6
+metric = "resolved"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "SWE-Bench Multilingual"
+score = 73.3
+metric = "resolved"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "BrowseComp"
+score = 73.2
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 45.1
+metric = "pass@1"
+variant = "preview checkpoint; max effort; with tools"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "MCP Atlas"
+score = 69
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+dataset = "public"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "GDPval-AA"
+score = 1395
+metric = "Elo"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
+
+[[benchmarks]]
+name = "Toolathlon"
+score = 47.8
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
diff --git a/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4-pro-0813.toml b/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4-pro-0813.toml
new file mode 100644
index 0000000000..dade234dbb
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4-pro-0813.toml
@@ -0,0 +1,106 @@
+# Public code-agent benchmarks use DeepSeek Harness minimal mode, max effort, temperature=1.0, top_p=0.95.
+# https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813
+name = "DeepSeek V4 Pro 0813"
+description = "DeepSeek V4 Pro snapshot with million-token context and support for thinking and non-thinking modes"
+family = "deepseek-thinking"
+release_date = "2026-08-12"
+last_updated = "2026-08-22"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+license = "MIT"
+
+[limit]
+context = 1_000_000
+output = 384_000
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 42.7
+metric = "score"
+variant = "without tools"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 60
+metric = "score"
+variant = "with tools"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 87.9
+metric = "score"
+harness = "DeepSeek Harness minimal mode"
+variant = "max effort"
+version = "2.1"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"
+
+[[benchmarks]]
+name = "NL2Repo"
+score = 61.5
+metric = "score"
+harness = "DeepSeek Harness minimal mode"
+variant = "max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"
+
+[[benchmarks]]
+name = "CyberGym"
+score = 83.3
+metric = "score"
+harness = "DeepSeek Harness minimal mode"
+variant = "max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 62.7
+metric = "score"
+harness = "DeepSeek Harness minimal mode"
+variant = "max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"
+
+[[benchmarks]]
+name = "Toolathlon-Verified"
+score = 74.1
+metric = "score"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"
+
+[[benchmarks]]
+name = "Agents' Last Exam"
+score = 25.7
+metric = "score"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"
+
+[[benchmarks]]
+name = "AutomationBench"
+score = 31.8
+metric = "score"
+dataset = "public"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"
+
+[[benchmarks]]
+name = "DSBench-FullStack"
+score = 71.1
+metric = "score"
+dataset = "internal"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"
+
+[[benchmarks]]
+name = "DSBench-Hard"
+score = 67.2
+metric = "score"
+dataset = "internal"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813"
diff --git a/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4-pro.toml b/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4-pro.toml
new file mode 100644
index 0000000000..c0bb4099ca
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4-pro.toml
@@ -0,0 +1,217 @@
+# Added benchmarks are from the original V4 preview instruct-model comparison, not the later 0731/0813 releases.
+name = "DeepSeek V4 Pro"
+description = "Open MoE flagship with million-token context for coding and long agent runs"
+family = "deepseek-thinking"
+release_date = "2026-04-24"
+last_updated = "2026-04-24"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2025-05"
+open_weights = true
+
+[limit]
+context = 1_000_000
+output = 384_000
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "SWE-Bench Verified"
+score = 80.6
+metric = "resolved"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "Artificial Analysis Coding Agent Index"
+score = 50.1
+metric = "average pass@1"
+harness = "Claude Code"
+variant = "high"
+source = "https://artificialanalysis.ai/agents/coding-agents"
+
+[[benchmarks]]
+name = "SWE-Atlas Codebase QnA"
+score = 67.8
+metric = "pass@1"
+harness = "Claude Code"
+variant = "high"
+source = "https://artificialanalysis.ai/agents/coding-agents"
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 18
+metric = "pass@1"
+harness = "Claude Code"
+variant = "high"
+dataset = "hard-aa"
+source = "https://artificialanalysis.ai/agents/coding-agents"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 64.7
+metric = "pass@1"
+harness = "Claude Code"
+variant = "high"
+version = "2.1"
+source = "https://artificialanalysis.ai/agents/coding-agents"
+
+[[benchmarks]]
+name = "MMLU-Pro"
+score = 87.5
+metric = "EM"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "SimpleQA-Verified"
+score = 57.9
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "Chinese SimpleQA"
+score = 84.4
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "GPQA Diamond"
+score = 90.1
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 37.7
+metric = "pass@1"
+variant = "preview checkpoint; max effort; without tools"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "LiveCodeBench"
+score = 93.5
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "Codeforces"
+score = 3206
+metric = "rating"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "HMMT"
+score = 95.2
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+dataset = "February 2026"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "IMOAnswerBench"
+score = 89.8
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "MathArena Apex"
+score = 38.3
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "MathArena Apex Shortlist"
+score = 90.2
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "MRCR"
+score = 83.5
+metric = "MMR"
+variant = "preview checkpoint; max effort"
+dataset = "1M context"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "CorpusQA"
+score = 62
+metric = "accuracy"
+variant = "preview checkpoint; max effort"
+dataset = "1M context"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 67.9
+metric = "accuracy"
+variant = "preview checkpoint; max effort"
+version = "2.0"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 55.4
+metric = "resolved"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "SWE-Bench Multilingual"
+score = 76.2
+metric = "resolved"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "BrowseComp"
+score = 83.4
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 48.2
+metric = "pass@1"
+variant = "preview checkpoint; max effort; with tools"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "MCP Atlas"
+score = 73.6
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+dataset = "public"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "GDPval-AA"
+score = 1554
+metric = "Elo"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
+
+[[benchmarks]]
+name = "Toolathlon"
+score = 51.8
+metric = "pass@1"
+variant = "preview checkpoint; max effort"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
diff --git a/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4.1-flash.toml b/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4.1-flash.toml
new file mode 100644
index 0000000000..16a1174cf7
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/deepseek/deepseek-v4.1-flash.toml
@@ -0,0 +1,311 @@
+# Benchmarks use instruct weights, temperature=1.0, top_p=0.95; not the base-model table.
+# Source: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash#instruct-model
+# Open weights (MIT): https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
+name = "DeepSeek V4.1 Flash"
+description = "DeepSeek V4.1 Flash model for reasoning and agentic coding"
+family = "deepseek-flash"
+release_date = "2026-09-10"
+last_updated = "2026-09-10"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2025-05"
+open_weights = true
+license = "MIT"
+
+[limit]
+context = 1_000_000
+output = 384_000
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "GPQA Diamond"
+score = 90.9
+metric = "pass@1"
+variant = "reasoning_effort=100"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "MathArena Apex"
+score = 65.6
+metric = "pass@1"
+variant = "reasoning_effort=100"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 36.8
+metric = "pass@1"
+variant = "reasoning_effort=100"
+dataset = "full set"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 39.1
+metric = "pass@1"
+variant = "reasoning_effort=100"
+dataset = "text-only subset"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Codeforces"
+score = 3471
+metric = "rating"
+variant = "reasoning_effort=100"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 90.6
+metric = "pass@1"
+variant = "reasoning_effort=100"
+version = "2.1"
+harness = "DeepSeek Harness minimal mode"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 30.0
+metric = "pass@1"
+variant = "reasoning_effort=100"
+version = "3.0"
+harness = "DeepSeek Harness minimal mode"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 31.2
+metric = "pass@1"
+variant = "reasoning_effort=100"
+version = "4.0"
+harness = "DeepSeek Harness minimal mode"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 74.2
+metric = "resolved"
+variant = "reasoning_effort=100"
+version = "1.1"
+harness = "mini-swe-agent"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "ProgramBench"
+score = 20.3
+metric = "almost@1"
+variant = "reasoning_effort=100"
+harness = "DeepSeek Harness minimal mode"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "NL2Repo"
+score = 64.0
+metric = "score"
+variant = "reasoning_effort=100"
+harness = "DeepSeek Harness minimal mode"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "CyberGym"
+score = 88.1
+metric = "pass@1"
+variant = "reasoning_effort=100"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "ExploitGym"
+score = 15.3
+metric = "pass@1"
+variant = "reasoning_effort=100"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "AutomationBench"
+score = 54.8
+metric = "pass@1"
+variant = "reasoning_effort=100"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Agents' Last Exam"
+score = 31.8
+metric = "pass@1"
+variant = "reasoning_effort=100"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "SEC-Bench Pro"
+score = 62.8
+metric = "pass@1"
+variant = "reasoning_effort=100"
+harness = "Claude Code"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 63.9
+metric = "pass@1"
+variant = "reasoning_effort=100; with tools"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Chartography"
+score = 78.9
+metric = "pass@1"
+variant = "reasoning_effort=100; with tools"
+harness = "Claude Code"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "BabyVision"
+score = 89.6
+metric = "pass@1"
+variant = "reasoning_effort=100; with tools"
+harness = "Claude Code"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "ZeroBench"
+score = 49.0
+metric = "pass@5"
+variant = "reasoning_effort=100; with tools"
+harness = "Claude Code"
+dataset = "main"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 69.8
+metric = "resolved"
+variant = "reasoning_effort=100"
+version = "1.1"
+harness = "Claude Code"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 88.0
+metric = "pass@1"
+variant = "reasoning_effort=100"
+version = "2.1"
+harness = "Claude Code"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 65.6
+metric = "resolved"
+variant = "reasoning_effort=100"
+version = "1.1"
+harness = "Codex"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 84.1
+metric = "pass@1"
+variant = "reasoning_effort=100"
+version = "2.1"
+harness = "Codex"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 65.5
+metric = "resolved"
+variant = "reasoning_effort=100"
+version = "1.1"
+harness = "OpenCode"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 85.0
+metric = "pass@1"
+variant = "reasoning_effort=100"
+version = "2.1"
+harness = "OpenCode"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 66.2
+metric = "resolved"
+variant = "reasoning_effort=100"
+version = "1.1"
+harness = "Pi"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 86.1
+metric = "pass@1"
+variant = "reasoning_effort=100"
+version = "2.1"
+harness = "Pi"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 90.3
+metric = "pass@1"
+variant = "reasoning_effort=100"
+version = "2.1"
+harness = "mini-swe-agent"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 72.6
+metric = "resolved"
+variant = "reasoning_effort=100"
+version = "1.1"
+harness = "DeepSeek Harness minimal mode"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 70.5
+metric = "resolved"
+variant = "reasoning_effort=100"
+version = "1.1"
+harness = "DeepSeek Harness standard mode"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 85.8
+metric = "pass@1"
+variant = "reasoning_effort=100"
+version = "2.1"
+harness = "DeepSeek Harness standard mode"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 67.6
+metric = "resolved"
+variant = "reasoning_effort=100"
+version = "1.1"
+harness = "DeepSeek Harness PTC mode"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 85.8
+metric = "pass@1"
+variant = "reasoning_effort=100"
+version = "2.1"
+harness = "DeepSeek Harness PTC mode"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash"
diff --git a/sdk/nexent/core/model_service/profiles/models/minimax/MiniMax-M2.7.toml b/sdk/nexent/core/model_service/profiles/models/minimax/MiniMax-M2.7.toml
new file mode 100644
index 0000000000..e72eaabdce
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/minimax/MiniMax-M2.7.toml
@@ -0,0 +1,46 @@
+name = "MiniMax-M2.7"
+description = "Open MiniMax flagship for coding agents, office automation, and complex environments"
+family = "minimax"
+release_date = "2026-03-18"
+last_updated = "2026-03-18"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = true
+
+[limit]
+context = 204_800
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/MiniMaxAI/MiniMax-M2.7"
+
+[[benchmarks]]
+name = "SWE-Bench Verified"
+score = 79.9
+metric = "resolved"
+harness = "Claude Code"
+source = "https://www.minimax.io/blog/minimax-m3"
+date = "2026-06-01"
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 56.2
+metric = "resolve rate"
+harness = "Claude Code"
+source = "https://www.minimax.io/blog/minimax-m3"
+date = "2026-06-01"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 51.1
+metric = "success rate"
+version = "2.1"
+source = "https://www.minimax.io/blog/minimax-m3"
+date = "2026-06-01"
diff --git a/sdk/nexent/core/model_service/profiles/models/minimax/MiniMax-M3.toml b/sdk/nexent/core/model_service/profiles/models/minimax/MiniMax-M3.toml
new file mode 100644
index 0000000000..31ad8eeb00
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/minimax/MiniMax-M3.toml
@@ -0,0 +1,67 @@
+name = "MiniMax-M3"
+description = "MiniMax multimodal model for long-context coding, perception, and agent planning"
+family = "minimax"
+release_date = "2026-06-01"
+last_updated = "2026-06-01"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = true
+
+[limit]
+context = 1_048_576
+output = 512_000
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/MiniMaxAI/MiniMax-M3"
+
+[[benchmarks]]
+name = "SWE-Bench Verified"
+score = 80.5
+metric = "resolved"
+harness = "Claude Code"
+source = "https://www.minimax.io/blog/minimax-m3"
+date = "2026-06-01"
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 59.0
+metric = "resolve rate"
+harness = "Claude Code"
+source = "https://www.minimax.io/blog/minimax-m3"
+date = "2026-06-01"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 66.0
+metric = "success rate"
+version = "2.1"
+source = "https://www.minimax.io/blog/minimax-m3"
+date = "2026-06-01"
+
+[[benchmarks]]
+name = "BrowseComp"
+score = 83.52
+metric = "accuracy"
+source = "https://www.minimax.io/blog/minimax-m3"
+date = "2026-06-01"
+
+[[benchmarks]]
+name = "MCP Atlas"
+score = 74.2
+metric = "success rate"
+source = "https://www.minimax.io/blog/minimax-m3"
+date = "2026-06-01"
+
+[[benchmarks]]
+name = "OSWorld-Verified"
+score = 70.06
+metric = "success rate"
+source = "https://www.minimax.io/blog/minimax-m3"
+date = "2026-06-01"
diff --git a/sdk/nexent/core/model_service/profiles/models/zhipuai/glm-5.1.toml b/sdk/nexent/core/model_service/profiles/models/zhipuai/glm-5.1.toml
new file mode 100644
index 0000000000..61860d5017
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/zhipuai/glm-5.1.toml
@@ -0,0 +1,53 @@
+name = "GLM-5.1"
+description = "Strong GLM coding model for agentic engineering, terminals, and repository generation"
+family = "glm"
+release_date = "2026-04-07"
+last_updated = "2026-04-07"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+
+[limit]
+context = 200_000
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/zai-org/GLM-5.1"
+
+[[benchmarks]]
+name = "Artificial Analysis Coding Agent Index"
+score = 52.7
+metric = "average pass@1"
+harness = "Claude Code"
+source = "https://artificialanalysis.ai/agents/coding-agents"
+
+[[benchmarks]]
+name = "SWE-Atlas Codebase QnA"
+score = 73.2
+metric = "pass@1"
+harness = "Claude Code"
+source = "https://artificialanalysis.ai/agents/coding-agents"
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 19.8
+metric = "pass@1"
+harness = "Claude Code"
+dataset = "hard-aa"
+source = "https://artificialanalysis.ai/agents/coding-agents"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 65.1
+metric = "pass@1"
+harness = "Claude Code"
+version = "2.1"
+source = "https://artificialanalysis.ai/agents/coding-agents"
diff --git a/sdk/nexent/core/model_service/profiles/models/zhipuai/glm-5.2.toml b/sdk/nexent/core/model_service/profiles/models/zhipuai/glm-5.2.toml
new file mode 100644
index 0000000000..292a8dff6d
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/zhipuai/glm-5.2.toml
@@ -0,0 +1,167 @@
+name = "GLM-5.2"
+description = "Open flagship GLM for long-horizon coding agents and million-token context work"
+family = "glm"
+release_date = "2026-06-13"
+last_updated = "2026-06-13"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+
+[limit]
+context = 1_000_000
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/zai-org/GLM-5.2"
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 62.1
+metric = "resolve rate"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 82.7
+metric = "success rate"
+harness = "Claude Code"
+version = "2.1"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "FrontierSWE"
+score = 74.4
+metric = "dominance"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 40.5
+metric = "accuracy"
+dataset = "text-only subset"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 54.7
+metric = "accuracy"
+variant = "with tools"
+dataset = "text-only subset"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "CritPt"
+score = 20.9
+metric = "accuracy"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "AIME"
+score = 99.2
+metric = "accuracy"
+version = "2026"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "HMMT"
+score = 94.4
+metric = "accuracy"
+version = "November 2025"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "HMMT"
+score = 92.5
+metric = "accuracy"
+version = "February 2026"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "IMOAnswerBench"
+score = 91.0
+metric = "accuracy"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "GPQA Diamond"
+score = 91.2
+metric = "accuracy"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "NL2Repo"
+score = 48.9
+metric = "resolve rate"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 46.2
+metric = "resolve rate"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "Program Bench"
+score = 63.7
+metric = "score"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 81.0
+metric = "success rate"
+harness = "Terminus 2"
+version = "2.1"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "PostTrainBench"
+score = 34.3
+metric = "score"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "SWE Marathon"
+score = 13.0
+metric = "resolve rate"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "MCP Atlas"
+score = 76.8
+metric = "score"
+dataset = "public subset"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "Tool-Decathlon"
+score = 48.2
+metric = "score"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
diff --git a/sdk/nexent/core/model_service/profiles/models/zhipuai/glm-5.3-flash.toml b/sdk/nexent/core/model_service/profiles/models/zhipuai/glm-5.3-flash.toml
new file mode 100644
index 0000000000..6f713eea56
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/zhipuai/glm-5.3-flash.toml
@@ -0,0 +1,70 @@
+# Scores: https://raw.githubusercontent.com/zai-org/GLM-5/refs/heads/main/resources/bench_53.png
+# Open weights: https://huggingface.co/zai-org/GLM-5.3-Flash (MIT, safetensors)
+name = "GLM-5.3-Flash"
+description = "Native multimodal GLM model for efficient coding and long-horizon agent tasks"
+family = "glm-flash"
+release_date = "2026-08-26"
+last_updated = "2026-08-26"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+
+[limit]
+context = 1_000_000
+output = 131_072
+
+[modalities]
+input = ["text", "image", "video", "pdf"]
+output = ["text"]
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 84.3
+metric = "pass@1"
+variant = "max effort"
+version = "2.1"
+harness = "Claude Code 2.1.207"
+source = "https://huggingface.co/zai-org/GLM-5.3-Flash"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 63.4
+metric = "resolved"
+variant = "max effort"
+version = "1.1"
+harness = "mini-swe-agent"
+source = "https://huggingface.co/zai-org/GLM-5.3-Flash"
+
+[[benchmarks]]
+name = "Agents' Last Exam"
+score = 26.3
+metric = "score"
+variant = "max effort"
+source = "https://huggingface.co/zai-org/GLM-5.3-Flash"
+
+[[benchmarks]]
+name = "AutomationBench"
+score = 48.8
+metric = "pass@1"
+variant = "max effort"
+version = "1.0.6"
+source = "https://huggingface.co/zai-org/GLM-5.3-Flash"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 55.3
+metric = "accuracy"
+variant = "max effort; with tools"
+dataset = "full set"
+source = "https://huggingface.co/zai-org/GLM-5.3-Flash"
+
+[[benchmarks]]
+name = "GDPval-AA"
+score = 1773
+metric = "Elo"
+variant = "max effort"
+version = "2"
+source = "https://huggingface.co/zai-org/GLM-5.3-Flash"
diff --git a/sdk/nexent/core/model_service/profiles/models/zhipuai/glm-5.3.toml b/sdk/nexent/core/model_service/profiles/models/zhipuai/glm-5.3.toml
new file mode 100644
index 0000000000..e8c11c17ae
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/models/zhipuai/glm-5.3.toml
@@ -0,0 +1,158 @@
+# Benchmark settings and caveats: https://huggingface.co/zai-org/GLM-5.3#footnotes
+name = "GLM-5.3"
+description = "Flagship GLM model for long-horizon coding, agents, and complex project delivery"
+family = "glm"
+release_date = "2026-08-14"
+last_updated = "2026-08-14"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+
+[limit]
+context = 1_000_000
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 88.2
+metric = "pass@1"
+variant = "max effort"
+version = "2.1"
+harness = "Claude Code 2.1.207"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 28.3
+metric = "avg@3"
+variant = "max effort"
+version = "3.0"
+harness = "Claude Code 2.1.207"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 66.9
+metric = "resolved"
+variant = "max effort; 6h timeout; 400K context"
+version = "1.1"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+harness = "mini-swe-agent"
+
+[[benchmarks]]
+name = "NL2Repo"
+score = 58.0
+metric = "score"
+variant = "max effort"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "ProgramBench"
+score = 19.0
+metric = "almost solved"
+variant = "max effort"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "FrontierSWE"
+score = 78.1
+metric = "dominance score"
+variant = "max effort"
+date = "2026-08-14"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "SWE-Marathon"
+score = 42.5
+metric = "score"
+variant = "max effort; modified anti-cheat checks"
+version = "1.1"
+harness = "Claude Code 2.1.207"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "PostTrainBench"
+score = 39.8
+metric = "weighted average"
+variant = "max effort; 3 runs; modified external-API checks"
+harness = "Claude Code 2.1.207"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "CyberGym"
+score = 84.5
+metric = "pass@1"
+variant = "max effort"
+harness = "Claude Code 2.1.207"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "ExploitGym"
+score = 105
+metric = "tasks solved"
+variant = "max effort; 2h rescaled timeout"
+dataset = "869 tasks"
+harness = "Claude Code 2.1.207"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "ExploitGym"
+score = 130
+metric = "tasks solved"
+variant = "max effort; 6h rescaled timeout"
+dataset = "869 tasks"
+harness = "Claude Code 2.1.207"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "ExploitBench"
+score = 54.4
+metric = "average coverage"
+variant = "max effort; union over 3 revisions"
+harness = "Claude Code 2.1.207"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "Toolathlon-Verified"
+score = 73
+metric = "pass@1"
+variant = "max effort"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "AutomationBench"
+score = 48.2
+metric = "pass@1"
+variant = "max effort"
+version = "1.0.6"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "Agents' Last Exam"
+score = 28.5
+metric = "score"
+variant = "max effort; CLI"
+harness = "Claude Code 2.1.207"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 62.5
+metric = "accuracy"
+variant = "max effort; with tools"
+source = "https://huggingface.co/zai-org/GLM-5.3"
+
+[[benchmarks]]
+name = "GDPval-AA"
+score = 1769
+metric = "Elo"
+variant = "max effort"
+version = "2"
+source = "https://huggingface.co/zai-org/GLM-5.3"
diff --git a/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/MiniMax/MiniMax-M2.7.toml b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/MiniMax/MiniMax-M2.7.toml
new file mode 100644
index 0000000000..80d471831f
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/MiniMax/MiniMax-M2.7.toml
@@ -0,0 +1,25 @@
+name = "MiniMax-M2.7"
+description = "MiniMax model for chat, coding, office work, and agentic tasks"
+family = "minimax"
+release_date = "2026-03-18"
+last_updated = "2026-03-18"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = true
+reasoning_options = []
+
+[cost]
+input = 0.30
+output = 1.20
+cache_read = 0.06
+cache_write = 0.375
+
+[limit]
+context = 204_800
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]
diff --git a/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/deepseek-v4-flash.toml b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/deepseek-v4-flash.toml
new file mode 100644
index 0000000000..f37c5ab26f
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/deepseek-v4-flash.toml
@@ -0,0 +1,16 @@
+base_model = "deepseek/deepseek-v4-flash"
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["high", "max"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.14
+output = 0.28
+cache_read = 0.0028
diff --git a/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/deepseek-v4-pro.toml b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/deepseek-v4-pro.toml
new file mode 100644
index 0000000000..722857a529
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/deepseek-v4-pro.toml
@@ -0,0 +1,16 @@
+base_model = "deepseek/deepseek-v4-pro"
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["high", "max"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.435
+output = 0.87
+cache_read = 0.003625
diff --git a/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/deepseek-v4.1-flash.toml b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/deepseek-v4.1-flash.toml
new file mode 100644
index 0000000000..6fe0ab8159
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/deepseek-v4.1-flash.toml
@@ -0,0 +1,27 @@
+# CNY→USD rate: 6.721845, 2026-09-18, source: https://open.er-api.com/v6/latest/USD
+# Sources (accessed 2026-09-18):
+# https://bailian.console.aliyun.com/cn-beijing?tab=model#/model-market/detail/deepseek-v4.1-flash?serviceSite=asia-pacific-china
+# https://help.aliyun.com/zh/model-studio/deepseek-api — capability table,
+# reasoning_effort ladder, max_tokens default (393_216, shared with thinking)
+# https://help.aliyun.com/zh/model-studio/billing-for-model-studio — Beijing
+# peak/off-peak pricing: CNY 2/1 input, 8/4 output per 1M tokens with context
+# cache discount; peak price converted to USD using the rate above (off-peak
+# is half).
+# Hybrid reasoning: enable_thinking = true|false toggles thinking;
+# reasoning_effort = low|high|max (default high) — low is only supported on
+# v4.1-flash / v4-flash-0731 / v4-pro-0813, and v4.1-flash is the current
+# low-tier bearer. reasoning_content streams in deltas. Responses API
+# supported (Beijing and Singapore only). Limits and modalities are identical
+# to the deepseek lab entry (1M context, 384_000 output, text+image input),
+# so they are inherited and not restated.
+base_model = "deepseek/deepseek-v4.1-flash"
+
+reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "high", "max"] }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.29754
+output = 1.19015
+cache_read = 0.01488
diff --git a/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/glm-5.1.toml b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/glm-5.1.toml
new file mode 100644
index 0000000000..e19b35f1e8
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/glm-5.1.toml
@@ -0,0 +1,45 @@
+# Chinese mainland deployment is tiered on input tokens per request; every
+# token in a request bills at its tier's unit price. The 0.87/3.48 previously
+# published here matches neither band; both bands below are taken from the
+# vendor page. No cache-read price is published for either band, so the
+# existing curated cache_read is retained on the base and not restated.
+# https://www.alibabacloud.com/help/en/model-studio/model-pricing (accessed 2026-08-26)
+name = "GLM-5.1"
+description = "Flagship GLM model for hybrid reasoning, coding, and agentic engineering"
+family = "glm"
+release_date = "2026-04-14"
+last_updated = "2026-04-14"
+attachment = false
+reasoning = true
+structured_output = true
+temperature = true
+tool_call = true
+open_weights = true
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "budget_tokens"
+max = 131_072
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.825
+output = 3.301
+cache_read = 0.17
+
+[[cost.tiers]]
+tier = { type = "context", size = 32_000 }
+input = 1.1
+output = 3.851
+
+[limit]
+context = 202_752
+output = 128_000
+
+[modalities]
+input = ["text"]
+output = ["text"]
diff --git a/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/glm-5.2.toml b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/glm-5.2.toml
new file mode 100644
index 0000000000..b847c801b2
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/glm-5.2.toml
@@ -0,0 +1,22 @@
+base_model = "zhipuai/glm-5.2"
+
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+# First party price citation here
+# https://bailian.console.alibabacloud.com/cn-beijing?tab=model#/model-market/detail/glm-5.2?serviceSite=asia-pacific-china
+[cost]
+input = 1.1
+output = 3.851
+cache_read = 0.275
+cache_write = 0
+
+[limit]
+context = 1000000
+output = 128_000
+
+[modalities]
+input = ["text"]
+output = ["text"]
diff --git a/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/glm-5.3.toml b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/glm-5.3.toml
new file mode 100644
index 0000000000..f6d0f6e67e
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/glm-5.3.toml
@@ -0,0 +1,25 @@
+# GLM-5.3 on Alibaba Bailian (Model Studio, Beijing): always-on thinking with
+# reasoning_effort low|high|max (default max). Thinking cannot be turned off
+# (enable_thinking only accepts true; passing false errors) and there is no
+# thinking_budget — so a single effort option, no toggle, no budget.
+# reasoning_content is streamed in deltas.
+# Pricing: Beijing CNY 8 / 28 / 2 (cache hit) per 1M tokens, same list price
+# as glm-5.2 on the same host (see providers/alibaba-cn/models/glm-5.2.toml).
+# Limits and modalities are identical to the zhipuai lab entry
+# (1M context, 131_072 output, text-only), so they are inherited and not restated.
+# https://bailian.console.aliyun.com/cn-beijing?tab=model#/model-market/detail/glm-5.3?serviceSite=asia-pacific-china
+# https://docs.bigmodel.cn/cn/guide/models/text/glm-5.3
+base_model = "zhipuai/glm-5.3"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "max"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 1.1
+output = 3.851
+cache_read = 0.275
+cache_write = 0
diff --git a/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.5-397b-a17b.toml b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.5-397b-a17b.toml
new file mode 100644
index 0000000000..8bafd8221f
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.5-397b-a17b.toml
@@ -0,0 +1,42 @@
+# Chinese mainland deployment is tiered on input tokens per request:
+# 0<Token<=128K bills at 0.172/1.032 and 128K<Token<=256K at 0.43/2.58, with
+# every token in a request billed at its tier's unit price. Thinking-mode and
+# non-thinking-mode output are priced the same in both bands.
+# https://www.alibabacloud.com/help/en/model-studio/model-pricing (accessed 2026-08-26)
+name = "Qwen3.5 397B-A17B"
+description = "Qwen vision-language model for visual reasoning, documents, and agent tasks"
+family = "qwen"
+release_date = "2026-02-16"
+last_updated = "2026-02-16"
+attachment = false
+reasoning = true
+temperature = true
+knowledge = "2025-04"
+tool_call = true
+open_weights = true
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "budget_tokens"
+max = 81_920
+
+[cost]
+input = 0.172
+output = 1.032
+reasoning = 1.032
+
+[[cost.tiers]]
+tier = { type = "context", size = 128_000 }
+input = 0.43
+output = 2.58
+reasoning = 2.58
+
+[limit]
+context = 262_144
+output = 65_536
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]
diff --git a/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.5-flash.toml b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.5-flash.toml
new file mode 100644
index 0000000000..4e8481bd07
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.5-flash.toml
@@ -0,0 +1,53 @@
+# Sources (accessed 2026-09-22):
+# https://help.aliyun.com/en/model-studio/model-pricing (China (Beijing) USD list prices)
+# https://help.aliyun.com/en/model-studio/qwen3-5-flash (capabilities and limits)
+# https://help.aliyun.com/en/model-studio/deep-thinking
+# Toggle: enable_thinking true|false
+# Budget: thinking_budget (integer reasoning tokens, max 81,920)
+name = "Qwen3.5 Flash"
+description = "Qwen vision-language model for visual reasoning, documents, and agent tasks"
+family = "qwen"
+release_date = "2026-02-23"
+last_updated = "2026-09-22"
+attachment = true
+reasoning = true
+structured_output = true
+temperature = true
+knowledge = "2025-04"
+tool_call = true
+open_weights = false
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "budget_tokens"
+max = 81_920
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.029
+output = 0.287
+reasoning = 0.287
+
+[[cost.tiers]]
+tier = { type = "context", size = 128_000 }
+input = 0.115
+output = 1.147
+reasoning = 1.147
+
+[[cost.tiers]]
+tier = { type = "context", size = 256_000 }
+input = 0.172
+output = 1.72
+reasoning = 1.72
+
+[limit]
+context = 1_000_000
+output = 65_536
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]
diff --git a/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.5-plus.toml b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.5-plus.toml
new file mode 100644
index 0000000000..429ae670c6
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.5-plus.toml
@@ -0,0 +1,53 @@
+# Sources (accessed 2026-09-22):
+# https://help.aliyun.com/en/model-studio/model-pricing (China (Beijing) USD list prices)
+# https://help.aliyun.com/en/model-studio/qwen3-5-plus (capabilities and limits)
+# https://help.aliyun.com/en/model-studio/deep-thinking
+# Toggle: enable_thinking true|false
+# Budget: thinking_budget (integer reasoning tokens, max 81,920)
+name = "Qwen3.5 Plus"
+description = "Qwen vision-language model for visual reasoning, documents, and agent tasks"
+family = "qwen"
+release_date = "2026-02-16"
+last_updated = "2026-09-22"
+attachment = true
+reasoning = true
+structured_output = true
+temperature = true
+knowledge = "2025-04"
+tool_call = true
+open_weights = false
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "budget_tokens"
+max = 81_920
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.115
+output = 0.688
+reasoning = 0.688
+
+[[cost.tiers]]
+tier = { type = "context", size = 128_000 }
+input = 0.287
+output = 1.72
+reasoning = 1.72
+
+[[cost.tiers]]
+tier = { type = "context", size = 256_000 }
+input = 0.573
+output = 3.44
+reasoning = 3.44
+
+[limit]
+context = 1_000_000
+output = 65_536
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]
diff --git a/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.8-flash.toml b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.8-flash.toml
new file mode 100644
index 0000000000..076549e30b
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.8-flash.toml
@@ -0,0 +1,26 @@
+# Sources (accessed 2026-08-28):
+# https://www.qianwenai.com/models/qwen3.8-flash
+# https://help.aliyun.com/zh/model-studio/model-pricing (Beijing: ¥0.8 input / ¥2.7
+# output per 1M tokens, flat 0<Token≤1M band, no tiers)
+# CNY→USD rate: 6.737174, 2026-08-28, source: https://open.er-api.com/v6/latest/USD
+# Cache pricing: cache read 10%, explicit cache create 125% of input
+# (https://help.aliyun.com/zh/model-studio/context-cache)
+# Toggle: enable_thinking true|false (hybrid)
+# Effort: reasoning_effort = low|medium|xhigh (not yet in the API reference; see
+# the Qwen3.8-Flash-Next release blog)
+# Budget: thinking_budget (Max Reasoning 262K)
+base_model = "alibaba/qwen3.8-flash"
+reasoning_options = [
+  { type = "toggle" },
+  { type = "effort", values = ["low", "medium", "xhigh"] },
+  { type = "budget_tokens", max = 262_144 },
+]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.11875
+output = 0.40073
+cache_read = 0.01187
+cache_write = 0.14844
diff --git a/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.8-max.toml b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.8-max.toml
new file mode 100644
index 0000000000..1f277d9455
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/models/qwen3.8-max.toml
@@ -0,0 +1,27 @@
+# Sources (accessed 2026-08-04):
+# https://www.qianwenai.com/models/qwen3.8-max
+# https://help.aliyun.com/zh/model-studio/model-pricing (Beijing: qwen3.8-max ¥12/¥36)
+# https://platform.qianwenai.com/docs/developer-guides/text-generation/thinking
+# CNY→USD rate: 6.7513, 2026-08-02, source: https://www.nimbuscalc.com/calculators/financial-calculators/currency-converter/USD-CNY-2026-08-02/
+# CNY list: input ¥12, output ¥36, implicit cache hit ¥1.5, explicit cache create ¥15,
+# explicit cache read ¥1 (per 1M tokens).
+# Toggle: enable_thinking true|false (hybrid)
+# Effort: reasoning_effort = low|medium|xhigh (default xhigh); high accepted as alias → xhigh
+# Budget: thinking_budget (0..262144) cannot be combined with reasoning_effort
+# Pay-as-you-go on DashScope CN / 千问AI (not Token Plan only).
+base_model = "alibaba/qwen3.8-max"
+structured_output = true
+reasoning_options = [
+  { type = "toggle" },
+  { type = "effort", values = ["low", "medium", "xhigh"] },
+  { type = "budget_tokens", min = 0, max = 262_144 },
+]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 1.77744
+output = 5.33231
+cache_read = 0.22218
+cache_write = 2.22179
diff --git a/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/provider.toml b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/provider.toml
new file mode 100644
index 0000000000..ff9e690f74
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/alibaba-cn/provider.toml
@@ -0,0 +1,20 @@
+name = "Alibaba (China)"
+env = ["DASHSCOPE_API_KEY"]
+npm = "@ai-sdk/openai-compatible"
+# Reasoning HTTP format (accessed 2026-06-25):
+# OpenAI Chat: POST /compatible-mode/v1/chat/completions; top-level
+# `enable_thinking` is true or false and `thinking_budget` is an integer token
+# cap (Qwen3 in thinking mode and Kimi only; default is the model maximum).
+# Responses: POST /compatible-mode/v1/responses; `reasoning.effort` is none,
+# minimal, low, medium (default), or high; no numeric thinking budget is accepted.
+# Anthropic: POST /apps/anthropic/v1/messages; `thinking.type` is enabled or
+# disabled and `thinking.budget_tokens` is an integer used only when enabled.
+# `reasoning_effort` is high or max only for DeepSeek V4 Pro/Flash.
+# DashScope text: POST /api/v1/services/aigc/text-generation/generation; put
+# `enable_thinking` and `thinking_budget` under `parameters`, not at JSON root.
+# Sources:
+# https://help.aliyun.com/zh/model-studio/deep-thinking
+# https://help.aliyun.com/zh/model-studio/compatibility-with-openai-responses-api
+# https://www.alibabacloud.com/help/en/model-studio/anthropic-api-messages
+doc = "https://www.alibabacloud.com/help/en/model-studio/models"
+api = "https://dashscope.aliyuncs.com/compatible-mode/v1" 
\ No newline at end of file
diff --git a/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/deepseek-v4-flash.toml b/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/deepseek-v4-flash.toml
new file mode 100644
index 0000000000..706a72594b
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/deepseek-v4-flash.toml
@@ -0,0 +1,23 @@
+# Coding Plan model name from https://www.volcengine.com/docs/82379/1928261
+# (accessed 2026-08-26). Subscription tier: no per-token price.
+# Toggle: thinking.type = enabled|disabled, POST /api/coding/v3/chat/completions
+# (verified: enabled -> reasoning_tokens 57, disabled -> 0).
+# No graded effort: the same body accepts reasoning_effort, but acceptance is not
+# effect. Over n=5 runs per level on a prompt hard enough to consume the budget,
+# none and minimal were hard zeros (the off switch) while the graded values fully
+# overlapped -- low 659/873/216/487/845, medium 271/1548/556/296/1245,
+# high 121/123/132/349/262, xhigh 928/753/268/480/110, max 811/435/570/910/623.
+# 'high' sits below 'low', so the levels are not separable and only the toggle is
+# declared, matching the toggle the lab (providers/deepseek) exposes.
+base_model = "deepseek/deepseek-v4-flash"
+
+[[reasoning_options]]
+type = "toggle"
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0
+output = 0
+cache_read = 0
diff --git a/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/deepseek-v4-pro.toml b/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/deepseek-v4-pro.toml
new file mode 100644
index 0000000000..68e7561f92
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/deepseek-v4-pro.toml
@@ -0,0 +1,22 @@
+# Coding Plan model name from https://www.volcengine.com/docs/82379/1928261
+# (accessed 2026-08-26). Subscription tier: no per-token price.
+# Toggle: thinking.type = enabled|disabled, POST /api/coding/v3/chat/completions
+# (verified: enabled -> reasoning_tokens 79, disabled -> 0).
+# No graded effort: the same body accepts reasoning_effort, but acceptance is not
+# effect. Over n=5 runs per level on a prompt hard enough to consume the budget,
+# none and minimal were hard zeros (the off switch) while the graded values fully
+# overlapped -- low 200/517/616/638/313, high 466/673/604/1075/415,
+# max 389/204/240/1535/688. The levels are not separable, so only the toggle is
+# declared, matching the toggle the lab (providers/deepseek) exposes.
+base_model = "deepseek/deepseek-v4-pro"
+
+[[reasoning_options]]
+type = "toggle"
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0
+output = 0
+cache_read = 0
diff --git a/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/glm-5.3-flash.toml b/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/glm-5.3-flash.toml
new file mode 100644
index 0000000000..45dd0b497e
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/glm-5.3-flash.toml
@@ -0,0 +1,17 @@
+# Coding Plan model name from https://www.volcengine.com/docs/82379/1928261
+# (accessed 2026-09-11): the model list includes both glm-5.3 (glm-latest)
+# and glm-5.3-flash. Subscription tier: no per-token price.
+# Effort levels mirror providers/zai-coding-plan/glm-5.3-flash.toml.
+base_model = "zhipuai/glm-5.3-flash"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "max"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0
+output = 0
+cache_read = 0
diff --git a/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/glm-5.3.toml b/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/glm-5.3.toml
new file mode 100644
index 0000000000..8d75263dbe
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/glm-5.3.toml
@@ -0,0 +1,30 @@
+# Coding Plan model name from https://www.volcengine.com/docs/82379/1928261
+# (accessed 2026-08-26). Subscription tier: no per-token price.
+# Effort: POST /api/coding/v3/chat/completions accepts low|medium|high|xhigh|max,
+# but only the three effective levels are declared. This model reports
+# reasoning_tokens = 0 on every level, so the measure is reasoning_content length;
+# over n=5 runs per level on a prompt hard enough to consume the thinking budget:
+#   low    0 0 0 0 0          medium 0 0 0 0 0            -> medium aliases low
+#   high   169 149 0 188 0
+#   xhigh  2840 1943 3338 4655 6279  max 2363 4567 2962 4606 2327
+#                                                          -> xhigh aliases max
+# Three distinct tiers, matching the lab (providers/zai) and the coding-plan peers.
+# Volcengine's mapping table (82379/1449737) covers glm-5-2-260617 but not 5.3, and
+# that row disagrees with what 5.3 does here (it maps low/medium to high, whereas
+# 5.3 returns no reasoning at all on both), so these levels are set from the
+# measurements above rather than from that table.
+# No toggle: both reasoning_effort=none|minimal and thinking.type=disabled are
+# rejected with 'not supported by this model'.
+base_model = "zhipuai/glm-5.3"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "max"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0
+output = 0
+cache_read = 0
diff --git a/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/minimax-m3.toml b/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/minimax-m3.toml
new file mode 100644
index 0000000000..60ca6e93c0
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/models/minimax-m3.toml
@@ -0,0 +1,21 @@
+# Coding Plan model name from https://www.volcengine.com/docs/82379/1928261
+# (accessed 2026-08-26). Subscription tier: no per-token price.
+# Toggle: thinking.type = enabled|disabled, POST /api/coding/v3/chat/completions
+# Toggle only, matching first-party MiniMax and minimax-coding-plan.
+# The same body also accepts reasoning_effort, but acceptance is not effect:
+# over n=3 runs per level on a hard arithmetic prompt, reasoning_content length was
+# minimal 1161/1033/805, high 1195/741/840, max 957/709/1088 -- fully overlapping,
+# so no graded control is declared. 'none' is likewise accepted without disabling
+# thinking (reasoning_tokens stays non-zero), so thinking.type is the real switch.
+base_model = "minimax/MiniMax-M3"
+
+[[reasoning_options]]
+type = "toggle"
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0
+output = 0
+cache_read = 0
diff --git a/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/provider.toml b/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/provider.toml
new file mode 100644
index 0000000000..7420f6ed56
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/volcengine-coding-plan/provider.toml
@@ -0,0 +1,42 @@
+name = "Volcengine Ark Coding Plan"
+env = ["ARK_CODING_PLAN_API_KEY"]
+npm = "@ai-sdk/openai-compatible"
+# Coding Plan is a subscription tier with its own base URL. Volcengine warns that
+# calling the pay-as-you-go https://ark.cn-beijing.volces.com/api/v3 with a Coding
+# Plan key does NOT draw down the subscription and bills separately, so the two
+# providers must stay distinct.
+# The env var is deliberately not ARK_API_KEY: /api/coding/v3 does not check which
+# plan a key belongs to, so a pay-as-you-go key reaches it and bills per token.
+# Sharing ARK_API_KEY would light up both providers at once and let a user pick a
+# Coding Plan entry while silently paying pay-as-you-go rates.
+# Reasoning HTTP format (accessed 2026-08-26):
+# OpenAI Chat: POST /api/coding/v3/chat/completions; `thinking.type` = enabled or
+# disabled (`auto` is rejected on most models); `reasoning_effort` is accepted on
+# every model but only some values have effect, so each file declares just the
+# levels it honours, measured on this endpoint (n=4-6 per level, 2026-08-27):
+#   doubao-seed-2.1-turbo / 2.0-lite / evolving -> none|low|high
+#       (medium overlaps low, xhigh and max overlap high rather than exceeding it)
+#   glm-5.3                                     -> low|high|max
+#       (rejects none and minimal outright; medium aliases low, xhigh aliases max)
+#   deepseek-v4-flash / -pro                    -> toggle only
+#       (graded values fully overlap; on Flash 'high' lands below 'low')
+#   kimi-k2.7-code                              -> no controls
+#   minimax-m3                                  -> toggle only
+#       (tolerates none but still spends reasoning tokens on it, so it is not an
+#        off switch there; thinking.type is)
+# Plaintext reasoning is in `choices[].message.reasoning_content`
+# and the encrypted block in `choices[].message.encrypted_content`.
+# Anthropic: base https://ark.cn-beijing.volces.com/api/coding, POST /v1/messages
+# with `thinking.type` and `output_config.effort`.
+# Model names here are the short console labels (doubao-seed-2.1-turbo). This
+# endpoint is the more permissive of the two: it accepts the dated IDs as well
+# (verified 2026-08-27 - doubao-seed-2-1-turbo-260628, doubao-seed-2-0-lite-260428
+# and glm-5-3-260930 all resolve here), whereas the pay-as-you-go /api/v3 accepts
+# only the dated form and 404s on a short label. The short labels are used because
+# they are what the Coding Plan console shows. `ark-code-latest` follows whichever
+# model is selected in the console.
+# Sources:
+# https://www.volcengine.com/docs/82379/1928261
+# https://www.volcengine.com/docs/82379/2546386
+doc = "https://www.volcengine.com/docs/82379/1928261"
+api = "https://ark.cn-beijing.volces.com/api/coding/v3"
diff --git a/sdk/nexent/core/model_service/profiles/providers/volcengine/models/deepseek-v4-flash-ga-260731.toml b/sdk/nexent/core/model_service/profiles/providers/volcengine/models/deepseek-v4-flash-ga-260731.toml
new file mode 100644
index 0000000000..85e6279908
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/volcengine/models/deepseek-v4-flash-ga-260731.toml
@@ -0,0 +1,21 @@
+# Model ID verified against POST /api/v3/chat/completions (2026-08-27): `deepseek-v4-flash-ga-260731`
+# CNY→USD rate: 6.737012, 2026-08-26, source: https://open.er-api.com/v6/latest/USD
+# CNY list price source: https://www.volcengine.com/docs/82379/1544106
+# Toggle: thinking.type = enabled|disabled
+# Effort: reasoning_effort = minimal|low|medium|high
+base_model = "deepseek/deepseek-v4-flash-0731"
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["minimal", "low", "medium", "high"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.4453
+output = 1.3359
+cache_read = 0.01484
diff --git a/sdk/nexent/core/model_service/profiles/providers/volcengine/models/deepseek-v4-pro-ga-260813.toml b/sdk/nexent/core/model_service/profiles/providers/volcengine/models/deepseek-v4-pro-ga-260813.toml
new file mode 100644
index 0000000000..369ff4ef40
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/volcengine/models/deepseek-v4-pro-ga-260813.toml
@@ -0,0 +1,21 @@
+# Model ID verified against POST /api/v3/chat/completions (2026-08-27): `deepseek-v4-pro-ga-260813`
+# CNY→USD rate: 6.737012, 2026-08-26, source: https://open.er-api.com/v6/latest/USD
+# CNY list price source: https://www.volcengine.com/docs/82379/1544106
+# Toggle: thinking.type = enabled|disabled
+# Effort: reasoning_effort = minimal|low|medium|high
+base_model = "deepseek/deepseek-v4-pro-0813"
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["minimal", "low", "medium", "high"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 1.3359
+output = 4.00771
+cache_read = 0.04453
diff --git a/sdk/nexent/core/model_service/profiles/providers/volcengine/models/glm-5-2-260617.toml b/sdk/nexent/core/model_service/profiles/providers/volcengine/models/glm-5-2-260617.toml
new file mode 100644
index 0000000000..64005aefe7
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/volcengine/models/glm-5-2-260617.toml
@@ -0,0 +1,21 @@
+# Model ID verified against POST /api/v3/chat/completions (2026-08-27): `glm-5-2-260617`
+# CNY→USD rate: 6.737012, 2026-08-26, source: https://open.er-api.com/v6/latest/USD
+# CNY list price source: https://www.volcengine.com/docs/82379/1544106
+# Toggle: thinking.type = enabled|disabled
+# Effort: reasoning_effort = minimal|low|medium|high
+base_model = "zhipuai/glm-5.2"
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["minimal", "low", "medium", "high"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 1.18747
+output = 4.15615
+cache_read = 0.29687
diff --git a/sdk/nexent/core/model_service/profiles/providers/volcengine/models/glm-5-3-flash-260828.toml b/sdk/nexent/core/model_service/profiles/providers/volcengine/models/glm-5-3-flash-260828.toml
new file mode 100644
index 0000000000..98cb87ee9b
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/volcengine/models/glm-5-3-flash-260828.toml
@@ -0,0 +1,18 @@
+# Model ID verified against POST /api/v3/chat/completions (2026-09-16): `glm-5-3-flash-260828`
+# CNY→USD rate: 6.737012, 2026-08-26, source: https://open.er-api.com/v6/latest/USD
+# CNY list price source: https://docs.volcengine.com/docs/82379/1449737 (深度思考) + 智谱官方价目表
+# GLM-5.3-Flash always reasons (thinking cannot be disabled on this model ID)
+# Effort: reasoning_effort = low|high|max
+base_model = "zhipuai/glm-5.3-flash"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "max"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.11875
+output = 0.41563
+cache_read = 0.03414
diff --git a/sdk/nexent/core/model_service/profiles/providers/volcengine/provider.toml b/sdk/nexent/core/model_service/profiles/providers/volcengine/provider.toml
new file mode 100644
index 0000000000..6ca805b6ab
--- /dev/null
+++ b/sdk/nexent/core/model_service/profiles/providers/volcengine/provider.toml
@@ -0,0 +1,23 @@
+name = "Volcengine Ark"
+env = ["ARK_API_KEY"]
+npm = "@ai-sdk/openai-compatible"
+# Reasoning HTTP format (accessed 2026-08-26):
+# OpenAI Chat: POST /api/v3/chat/completions; `thinking.type` = enabled|disabled
+# (`auto` is documented but rejected with "Unsupported thinking type for the
+# current model" on every model here except doubao-seed-1-6-flash),
+# `reasoning_effort` = none|minimal|low|medium|high|xhigh|max (none and minimal
+# disable thinking; doubao-seed-1-6-251015 accepts only the middle four).
+# Plaintext reasoning is in
+# `choices[].message.reasoning_content`, encrypted block is in
+# `choices[].message.encrypted_content` (returned by default).
+# Responses: POST /api/v3/responses; `reasoning.effort`; encrypted block requires
+# explicit `include = ["reasoning.encrypted_content"]`.
+# Anthropic: POST /api/v3/compatible/v1/messages; `thinking.type`,
+# `output_config.effort` = low|high|max; plaintext thinking is in
+# `content[thinking].thinking`, signature in `content[thinking].signature` (returned by default).
+# Sources:
+# https://www.volcengine.com/docs/82379/1298459
+# https://www.volcengine.com/docs/82379/1330310
+# https://www.volcengine.com/docs/82379/2636748
+doc = "https://www.volcengine.com/docs/82379/1330310"
+api = "https://ark.cn-beijing.volces.com/api/v3"
diff --git a/sdk/pyproject.toml b/sdk/pyproject.toml
index 4f95a9ac31..b1f928ca20 100644
--- a/sdk/pyproject.toml
+++ b/sdk/pyproject.toml
@@ -103,6 +103,7 @@ exclude = ["tests*", "examples*"]
 namespaces = true
 
 [tool.setuptools.package-data]
+"nexent.core.model_service.profiles" = ["**/*.toml", "README.md", "LICENSE"]
 "nexent.core.prompts" = ["zh/*/*.yaml", "en/*/*.yaml"]
 "nexent.assets" = ["*.png", "*.jpg", "*.gif", "*.webp"]
 
diff --git a/test-e2e/changes/requirements/MODEL-PROFILE-CATALOG-20261013.yaml b/test-e2e/changes/requirements/MODEL-PROFILE-CATALOG-20261013.yaml
new file mode 100644
index 0000000000..26c17cd90e
--- /dev/null
+++ b/test-e2e/changes/requirements/MODEL-PROFILE-CATALOG-20261013.yaml
@@ -0,0 +1,43 @@
+schema_version: '1.0'
+change:
+  change_id: MODEL-PROFILE-CATALOG-20261013
+  change_type: requirement
+  requirement_id: MODEL-PROFILE-CATALOG-20261013
+  title: Package the pinned China-region model profile source inventory
+  description: Catalog source slice; no runtime consumer or support certification
+    is introduced.
+  affected_features:
+    existing: []
+    added:
+    - SDK-MODEL-PROFILE-CATALOG
+    modified: []
+    retired: []
+  affected_cases:
+    existing: []
+    added: []
+    modified: []
+    retired: []
+  traditional_ut:
+    decision: add
+    behaviors:
+    - The reviewed inventory has 21 model, 21 provider-model and 3 provider files,
+      all with verified manifest hashes.
+    - Every declared base_model dependency is in the selection; TOML parses without
+      requiring upstream TypeScript.
+    - SDK wheels retain the manifest, source files, README and complete upstream MIT
+      license.
+    - Controlled TOML reader rejects path escapes, invalid manifest versions, duplicates,
+      missing sources, digest mismatch and malformed TOML without returning a partial
+      inventory.
+    - Read source documents preserve exact source identity and deeply immutable nested
+      data; loading is not semantic normalization.
+    reason: Isolated source/package integrity checks; later reader, normalization
+      and composition steps need their own component verification.
+    test_selectors:
+    - test/sdk/core/model_service/test_catalog_inventory.py
+    - test/sdk/core/model_service/test_source_reader.py
+  no_case_change_reason: No API or runtime behavior is connected. Existing Cases remain
+    unchanged; installed-package inspection supplies artifact evidence, not Agent
+    acceptance.
+  notes: SPEC AC-01/16 subset. Source inventory is distinct from verified support;
+    live-model acceptance remains pending.
diff --git a/test-e2e/changes/requirements/MODEL-PROFILE-CONTRACTS-20261013.yaml b/test-e2e/changes/requirements/MODEL-PROFILE-CONTRACTS-20261013.yaml
new file mode 100644
index 0000000000..e260eb7bf3
--- /dev/null
+++ b/test-e2e/changes/requirements/MODEL-PROFILE-CONTRACTS-20261013.yaml
@@ -0,0 +1,28 @@
+schema_version: '1.0'
+change:
+  change_id: MODEL-PROFILE-CONTRACTS-20261013
+  change_type: requirement
+  requirement_id: MODEL-PROFILE-CONTRACTS-20261013
+  title: Add strict model profile contract primitives
+  description: First local implementation slice of the approved model-access / NativeAgent SPEC; no existing runtime is connected or modified.
+  affected_features:
+    existing: []
+    added: [SDK-MODEL-PROFILE-CONTRACTS]
+    modified: []
+    retired: []
+  affected_cases:
+    existing: []
+    added: []
+    modified: []
+    retired: []
+  traditional_ut:
+    decision: add
+    behaviors:
+      - Unknown capacity remains null and known limits reject booleans, numeric strings, zero, negatives and service limits above the combined window.
+      - Toggle, effort and budget controls preserve explicitly declared values and reject duplicate modes/keys, invalid types, ranges, unknown fields and unsupported-with-controls.
+      - Source references retain exact identity/digests, reject unsafe paths and round-trip through strict JSON validation without loading external sources.
+    reason: These are isolated domain validation contracts; existing model UI/runtime Cases do not exercise them, and a duplicate formal D1 would add no component evidence. Catalog parsing/composition and live runtime verification belong to subsequent slices.
+    test_selectors:
+      - test/sdk/core/model_service/test_profile_contracts.py
+  no_case_change_reason: No API, provider, database, tool execution or user journey is connected in this slice. Preserve all historical Cases; formal integration coverage will be added with the owning implementation.
+  notes: Implements only profile primitives under SPEC AC-01/02. Overall feature acceptance remains pending; neither formal runtime acceptance nor model compatibility is claimed.
diff --git a/test-e2e/features/SDK-MODEL-PROFILE-CATALOG/cases.md b/test-e2e/features/SDK-MODEL-PROFILE-CATALOG/cases.md
new file mode 100644
index 0000000000..a1a090f4ea
--- /dev/null
+++ b/test-e2e/features/SDK-MODEL-PROFILE-CATALOG/cases.md
@@ -0,0 +1,5 @@
+# SDK-MODEL-PROFILE-CATALOG 关联用例
+
+此页由结构化用例生成，不手工编辑。
+
+当前测试基线未保留关联用例。
diff --git a/test-e2e/features/SDK-MODEL-PROFILE-CATALOG/feature.yaml b/test-e2e/features/SDK-MODEL-PROFILE-CATALOG/feature.yaml
new file mode 100644
index 0000000000..899f55e1fe
--- /dev/null
+++ b/test-e2e/features/SDK-MODEL-PROFILE-CATALOG/feature.yaml
@@ -0,0 +1,22 @@
+schema_version: '1.0'
+module:
+  id: SDK-MODEL-SERVICE
+  name: SDK model service
+feature:
+  feature_id: SDK-MODEL-PROFILE-CATALOG
+  name: Local model profile source catalog
+  status: planned
+  description: Pinned source inventory and package resources for the upcoming local
+    resolver.
+  business_rules:
+  - rule_id: SDK-MODEL-PROFILE-CATALOG-RULE-01
+    description: Selected authored TOML retains exact upstream bytes and complete
+      base_model dependencies.
+  - rule_id: SDK-MODEL-PROFILE-CATALOG-RULE-02
+    description: Inventory inclusion is not verified model support; runtime uses saved
+      resolved outlets.
+  required_test_stages: []
+  source:
+    requirement_ids:
+    - MODEL-PROFILE-CATALOG-20261013
+    notes: Source and package integrity only; no runtime integration.
diff --git a/test-e2e/features/SDK-MODEL-PROFILE-CONTRACTS/cases.md b/test-e2e/features/SDK-MODEL-PROFILE-CONTRACTS/cases.md
new file mode 100644
index 0000000000..903c7086ec
--- /dev/null
+++ b/test-e2e/features/SDK-MODEL-PROFILE-CONTRACTS/cases.md
@@ -0,0 +1,5 @@
+# SDK-MODEL-PROFILE-CONTRACTS 关联用例
+
+此页由结构化用例生成，不手工编辑。
+
+当前测试基线未保留关联用例。
diff --git a/test-e2e/features/SDK-MODEL-PROFILE-CONTRACTS/feature.yaml b/test-e2e/features/SDK-MODEL-PROFILE-CONTRACTS/feature.yaml
new file mode 100644
index 0000000000..5b6a238185
--- /dev/null
+++ b/test-e2e/features/SDK-MODEL-PROFILE-CONTRACTS/feature.yaml
@@ -0,0 +1,20 @@
+schema_version: '1.0'
+module:
+  id: SDK-MODEL-SERVICE
+  name: SDK model service
+feature:
+  feature_id: SDK-MODEL-PROFILE-CONTRACTS
+  name: Model profile contract primitives
+  status: planned
+  description: Strict provider-neutral capacity, reasoning and source-reference contracts used by the upcoming catalog and outlet layers.
+  business_rules:
+    - rule_id: SDK-MODEL-PROFILE-CONTRACTS-RULE-01
+      description: Unknown capacity and capability stay distinct from explicit unsupported; known token limits are positive strict integers.
+    - rule_id: SDK-MODEL-PROFILE-CONTRACTS-RULE-02
+      description: Reasoning controls use declared toggle, effort or token-budget domains; duplicates, invalid types, inconsistent support and unknown fields fail validation.
+    - rule_id: SDK-MODEL-PROFILE-CONTRACTS-RULE-03
+      description: Atomic references carry exact source identity and digests with relative source paths; they are traceability, not runtime recomposition instructions.
+  required_test_stages: []
+  source:
+    requirement_ids: [MODEL-PROFILE-CONTRACTS-20261013]
+    notes: SPEC AC-01 and AC-02 subset only; the current slice has no catalog, persistence, provider or Agent runtime integration. Traditional UT proves these domain contracts. Later slices supply independent component/API/runtime evidence.
diff --git a/test/sdk/core/model_service/test_catalog_inventory.py b/test/sdk/core/model_service/test_catalog_inventory.py
new file mode 100644
index 0000000000..b9b23103d8
--- /dev/null
+++ b/test/sdk/core/model_service/test_catalog_inventory.py
@@ -0,0 +1,60 @@
+"""AC-01/16: integrity of the reviewed source inventory, not model support."""
+
+import hashlib
+import tomllib
+from collections import Counter
+from pathlib import Path
+
+import pytest
+
+CATALOG_ROOT = (
+    Path(__file__).resolve().parents[4] / "sdk/nexent/core/model_service/profiles"
+)
+MANIFEST = tomllib.loads((CATALOG_ROOT / "manifest.toml").read_text())
+
+
+def test_reviewed_inventory_is_pinned_and_complete():
+    assert MANIFEST["schema_version"] == 1
+    assert MANIFEST["source_schema_ref"] == "models_dev_authored_v1"
+    assert MANIFEST["upstream_revision"] == "e69e1844210e5cfe7c7e7f68d57328bc7451a51c"
+    sources = MANIFEST["sources"]
+    assert Counter(source["kind"] for source in sources) == {
+        "model": 21,
+        "provider_model": 21,
+        "provider": 3,
+    }
+    paths = [source["path"] for source in sources]
+    assert len(paths) == len(set(paths))
+    actual = {
+        str(path.relative_to(CATALOG_ROOT))
+        for folder in ("models", "providers")
+        for path in (CATALOG_ROOT / folder).rglob("*.toml")
+    }
+    assert actual == set(paths)
+
+
+@pytest.mark.parametrize(
+    "source", MANIFEST["sources"], ids=lambda source: source["path"]
+)
+def test_source_bytes_match_manifest_and_dependencies_are_present(source):
+    path = CATALOG_ROOT / source["path"]
+    data = path.read_bytes()
+    assert hashlib.sha256(data).hexdigest() == source["content_sha256"]
+    parsed = tomllib.loads(data.decode("utf-8"))
+    if "base_model" in parsed:
+        assert f"models/{parsed['base_model']}.toml" in {
+            item["path"] for item in MANIFEST["sources"] if item["kind"] == "model"
+        }
+    if source["kind"] == "provider":
+        assert parsed["api"].startswith("https://")
+        assert isinstance(parsed["env"], list)
+
+
+def test_license_and_scope_notice_ship_with_inventory():
+    license_text = (CATALOG_ROOT / "LICENSE").read_text()
+    assert "Copyright (c) 2025 models.dev" in license_text
+    assert "Permission is hereby granted, free of charge" in license_text
+    assert "THE SOFTWARE IS PROVIDED" in license_text
+    readme = (CATALOG_ROOT / "README.md").read_text()
+    assert MANIFEST["upstream_revision"] in readme
+    assert "not a verified support list" in readme
diff --git a/test/sdk/core/model_service/test_profile_contracts.py b/test/sdk/core/model_service/test_profile_contracts.py
new file mode 100644
index 0000000000..0bd9ee1c7b
--- /dev/null
+++ b/test/sdk/core/model_service/test_profile_contracts.py
@@ -0,0 +1,184 @@
+"""AC-01/02: strict profile primitives, without provider or database I/O."""
+
+import pytest
+from nexent.core.model_service.profiles.contracts import (
+    AtomicProfileReference,
+    BudgetTokensControl,
+    ContextCapabilities,
+    EffortControl,
+    ReasoningCapabilities,
+    ToggleControl,
+)
+from pydantic import ValidationError
+
+
+def test_context_unknown_and_independent_limits():
+    assert ContextCapabilities().model_dump() == {
+        "combined_window": None,
+        "input_limit": None,
+        "output_limit": None,
+    }
+    profile = ContextCapabilities(
+        combined_window=100, input_limit=100, output_limit=100
+    )
+    assert ContextCapabilities.model_validate_json(profile.model_dump_json()) == profile
+    assert ContextCapabilities(output_limit=1).combined_window is None
+
+
+@pytest.mark.parametrize("value", [True, False, "200000", 1.0, 0, -1])
+@pytest.mark.parametrize("field", ["combined_window", "input_limit", "output_limit"])
+def test_context_rejects_invalid_token_limits(field, value):
+    with pytest.raises(ValidationError):
+        ContextCapabilities(**{field: value})
+
+
+@pytest.mark.parametrize("field", ["input_limit", "output_limit"])
+def test_context_rejects_service_limit_above_window(field):
+    with pytest.raises(ValidationError):
+        ContextCapabilities(combined_window=100, **{field: 101})
+
+
+def test_reasoning_controls_preserve_declared_values_and_round_trip():
+    profile = ReasoningCapabilities(
+        status="supported",
+        controls=[
+            ToggleControl(values=[False, True]),
+            EffortControl(values=["low", "xhigh", "provider-specific"]),
+            BudgetTokensControl(min=0, max=100, special_values=[-1]),
+        ],
+        response_paths=["choices[].delta.reasoning_content"],
+    )
+    assert (
+        ReasoningCapabilities.model_validate_json(profile.model_dump_json()) == profile
+    )
+    assert profile.controls[0].values[0] is False
+    assert profile.controls[2].special_values == [-1]
+    assert "current_value" not in profile.model_dump_json()
+
+
+def test_reasoning_unknown_and_explicit_no_controls_remain_distinct():
+    assert ReasoningCapabilities().controls is None
+    assert ReasoningCapabilities(status="unsupported", controls=[]).controls == []
+    assert (
+        ReasoningCapabilities(controls=[ToggleControl(values=[True])]).status
+        == "unknown"
+    )
+
+
+@pytest.mark.parametrize(
+    "control",
+    [
+        {"mode": "toggle", "values": [1]},
+        {"mode": "toggle", "values": [False, False]},
+        {"mode": "toggle", "values": []},
+        {"mode": "toggle", "values": [True], "parameter_key": "reasoning"},
+        {"mode": "effort", "values": ["high", "high"]},
+        {"mode": "effort", "values": [" "]},
+        {"mode": "effort", "values": [1]},
+        {"mode": "effort", "values": []},
+        {"mode": "budget_tokens", "min": 10, "max": 9},
+        {"mode": "budget_tokens", "min": -1, "max": 9},
+        {"mode": "budget_tokens", "min": 0, "max": True},
+        {"mode": "budget_tokens", "min": 0, "max": 9, "special_values": [False]},
+        {"mode": "budget_tokens", "min": 0, "max": 9, "special_values": [-1, -1]},
+        {"mode": "unknown"},
+        {"mode": "toggle", "values": [True], "unexpected": 1},
+    ],
+)
+def test_reasoning_rejects_invalid_controls(control):
+    with pytest.raises(ValidationError):
+        ReasoningCapabilities(controls=[control])
+
+
+@pytest.mark.parametrize(
+    "payload",
+    [
+        {"status": "unsupported", "controls": [{"mode": "toggle", "values": [True]}]},
+        {"status": "unsupported", "response_paths": ["reasoning_content"]},
+        {
+            "status": "supported",
+            "controls": [
+                {"mode": "effort", "values": ["low"]},
+                {"mode": "effort", "values": ["high"]},
+            ],
+        },
+        {"response_paths": ["reasoning_content", "reasoning_content"]},
+        {"response_paths": [""]},
+        {"status": True},
+        {"status": "unknown", "extra": 1},
+    ],
+)
+def test_reasoning_rejects_conflicting_or_unknown_fields(payload):
+    with pytest.raises(ValidationError):
+        ReasoningCapabilities(**payload)
+
+
+def test_reasoning_size_limit_counts_utf8_bytes():
+    ReasoningCapabilities(controls=[EffortControl(values=["a" * 65000])])
+    with pytest.raises(ValidationError, match="64 KiB"):
+        ReasoningCapabilities(controls=[EffortControl(values=["思" * 22000])])
+
+
+def reference_payload():
+    return {
+        "key": "qwen/Qwen3.8",
+        "catalog_revision": "catalog-v1",
+        "file_path": "models/qwen/model.toml",
+        "content_sha256": "a" * 64,
+    }
+
+
+def test_atomic_reference_round_trip_preserves_exact_identity():
+    profile = AtomicProfileReference(**reference_payload())
+    assert (
+        AtomicProfileReference.model_validate_json(profile.model_dump_json()) == profile
+    )
+    assert profile.key == "qwen/Qwen3.8"
+
+
+@pytest.mark.parametrize(
+    "path",
+    [
+        "/tmp/model.toml",
+        "../x.toml",
+        "models/../x.toml",
+        "models/./x.toml",
+        "models//x.toml",
+        "C:/x.toml",
+        "models\\x.toml",
+        "",
+        "models/",
+        "models/\x00file.toml",
+    ],
+)
+def test_atomic_reference_rejects_unsafe_paths(path):
+    with pytest.raises(ValidationError):
+        AtomicProfileReference(**{**reference_payload(), "file_path": path})
+
+
+@pytest.mark.parametrize(
+    "updates",
+    [
+        {"content_sha256": "a" * 63},
+        {"content_sha256": "A" * 64},
+        {"content_sha256": "g" * 64},
+        {"key": " "},
+        {"catalog_revision": ""},
+        {"unknown": "field"},
+    ],
+)
+def test_atomic_reference_rejects_invalid_provenance(updates):
+    with pytest.raises(ValidationError):
+        AtomicProfileReference(**{**reference_payload(), **updates})
+
+
+def test_context_rejects_unknown_fields():
+    with pytest.raises(ValidationError):
+        ContextCapabilities(max_tokens=100)
+
+
+def test_contract_fields_cannot_be_reassigned_without_validation():
+    profile = ContextCapabilities(combined_window=100)
+    with pytest.raises(ValidationError):
+        profile.combined_window = 0
+    assert profile.combined_window == 100
diff --git a/test/sdk/core/model_service/test_source_reader.py b/test/sdk/core/model_service/test_source_reader.py
new file mode 100644
index 0000000000..064d849e1f
--- /dev/null
+++ b/test/sdk/core/model_service/test_source_reader.py
@@ -0,0 +1,134 @@
+"""Unit proof for controlled local sources; normalization is a separate contract."""
+
+import hashlib
+from pathlib import Path
+
+import pytest
+from nexent.core.model_service.profiles._models_dev.reader import (
+    SourceReadError,
+    read_sources,
+)
+
+CATALOG_ROOT = (
+    Path(__file__).resolve().parents[4] / "sdk/nexent/core/model_service/profiles"
+)
+
+
+def write_inventory(
+    root,
+    *,
+    data=b'name = "Test"\n[limit]\ncontext = 100\n',
+    path="models/test/model.toml",
+    kind="model",
+    digest=None,
+    version="1",
+    extra="",
+    copies=1,
+):
+    digest = digest or hashlib.sha256(data).hexdigest()
+    (root / "manifest.toml").write_text(
+        f'schema_version = {version}\nsource_schema_ref = "models_dev_authored_v1"\n'
+        'upstream_url = "https://github.com/anomalyco/models.dev"\n'
+        f'upstream_revision = "{"a" * 40}"\n{extra}\n'
+        + (
+            f'[[sources]]\nkind = "{kind}"\npath = "{path}"\ncontent_sha256 = "{digest}"\n'
+            * copies
+        )
+    )
+    if not path.startswith("/") and ".." not in path.split("/"):
+        target = root / path
+        target.parent.mkdir(parents=True, exist_ok=True)
+        target.write_bytes(data)
+
+
+def test_loads_pinned_inventory_with_exact_nested_model_identity():
+    documents = read_sources(CATALOG_ROOT)
+    assert len(documents) == 45
+    assert any(
+        document.key == "alibaba-cn/MiniMax/MiniMax-M2.7" for document in documents
+    )
+    assert {document.key for document in documents if document.kind == "provider"} == {
+        "alibaba-cn",
+        "volcengine",
+        "volcengine-coding-plan",
+    }
+
+
+def test_authored_documents_are_deeply_immutable(tmp_path):
+    write_inventory(
+        tmp_path,
+        data=b'name = "Test"\nvalues = ["low", "high"]\n[limit]\ncontext = 100\n',
+    )
+    documents = read_sources(tmp_path)
+    assert isinstance(documents, tuple)
+    doc = documents[0]
+    assert doc.data["values"] == ("low", "high")
+    with pytest.raises(TypeError):
+        doc.data["limit"]["context"] = 0
+    with pytest.raises(TypeError):
+        doc.data["name"] = "Changed"
+
+
+@pytest.mark.parametrize(
+    "changes,category",
+    [
+        ({"path": "../outside.toml"}, "SOURCE_REFERENCE_INVALID"),
+        ({"path": "/tmp/outside.toml"}, "SOURCE_REFERENCE_INVALID"),
+        ({"path": "models/./test.toml"}, "SOURCE_REFERENCE_INVALID"),
+        (
+            {"path": "providers/test/provider.toml", "kind": "model"},
+            "SOURCE_KIND_PATH_MISMATCH",
+        ),
+        ({"digest": "a" * 64}, "SOURCE_DIGEST_MISMATCH"),
+        ({"digest": "wrong"}, "SOURCE_REFERENCE_INVALID"),
+        ({"data": b"name = ["}, "SOURCE_TOML_INVALID"),
+        ({"data": b"\xff"}, "SOURCE_TOML_INVALID"),
+        ({"copies": 2}, "SOURCE_DUPLICATE"),
+        ({"copies": 0}, "SOURCE_MANIFEST_INVALID"),
+        ({"version": "2"}, "SOURCE_MANIFEST_INVALID"),
+        ({"version": "1.0"}, "SOURCE_MANIFEST_INVALID"),
+        ({"path": "models/test/model.json"}, "SOURCE_KIND_PATH_MISMATCH"),
+        ({"path": "models/test/.toml"}, "SOURCE_KIND_PATH_MISMATCH"),
+        ({"version": "true"}, "SOURCE_MANIFEST_INVALID"),
+        ({"extra": "unexpected = 1"}, "SOURCE_MANIFEST_INVALID"),
+    ],
+)
+def test_rejects_invalid_inventory_before_returning_partial_data(
+    tmp_path, changes, category
+):
+    write_inventory(tmp_path, **changes)
+    with pytest.raises(SourceReadError) as caught:
+        read_sources(tmp_path)
+    assert caught.value.category == category
+    assert "name =" not in str(caught.value)
+    assert caught.value.rule_id == "source_inventory_v1"
+
+
+def test_missing_manifest_and_source_are_explicit_errors(tmp_path):
+    with pytest.raises(SourceReadError, match="SOURCE_FILE_UNAVAILABLE"):
+        read_sources(tmp_path)
+    write_inventory(tmp_path)
+    (tmp_path / "models/test/model.toml").unlink()
+    with pytest.raises(SourceReadError, match="SOURCE_FILE_UNAVAILABLE"):
+        read_sources(tmp_path)
+
+
+def test_symlink_to_outside_root_is_rejected(tmp_path):
+    root = tmp_path / "catalog"
+    root.mkdir()
+    write_inventory(root)
+    target = root / "models/test/model.toml"
+    target.unlink()
+    outside = tmp_path / "outside.toml"
+    outside.write_text('name = "Outside"')
+    target.symlink_to(outside)
+    with pytest.raises(SourceReadError, match="SOURCE_PATH_OUTSIDE_ROOT"):
+        read_sources(root)
+
+
+def test_invalid_upstream_commit_is_rejected(tmp_path):
+    write_inventory(tmp_path)
+    manifest = tmp_path / "manifest.toml"
+    manifest.write_text(manifest.read_text().replace("a" * 40, "not-a-commit"))
+    with pytest.raises(SourceReadError, match="SOURCE_MANIFEST_INVALID"):
+        read_sources(tmp_path)