the AI companion that's actually yours.
she talks out loud, listens in real time, does things for you — and remembers your life together. her memory lives on your machine; her mind is NVIDIA Nemotron on Nebius Token Factory — or a local model, if you'd rather nothing leave it at all.
Experimental preview. Windows-first; expect rough edges.
People who spend long hours alone at their computer — students, remote workers, solo builders — who want an always-on companion they can simply talk to: one that remembers their life and helps with everyday tasks, without their private memories living on someone else's server.
Windows (PowerShell):
iex (irm https://yumii.me/install.ps1)A short setup asks you to pick her mind, ears and voice — the recommended mind is NVIDIA Nemotron on Nebius Token Factory, which needs one API key from tokenfactory.nebius.com. She lands in your Start Menu, downloads her voice the first time you open her, and then you just talk.
Updating: re-run the same command.
git clone https://github.com/CodeNeuron58/Yumii && cd Yumii
uv sync --extra dev # Python backend (needs uv)
cd desktop && npx @tauri-apps/cli dev # the desktop app (needs Rust + Node)Paste your Nebius Token Factory key in the first-run card (or the dashboard). Optional: a Tavily key in the dashboard gives her Tavily-powered web search.
uv tool uninstall yumii # the backend
Remove-Item "$env:LOCALAPPDATA\Yumii" -Recurse # the app
Remove-Item "$env:APPDATA\Microsoft\Windows\Start Menu\Programs\Yumii.lnk"Delete ~/.yumii too and she forgets everything — keep it and she's
portable.
- talks in real time — streamed speech, barge-in, and end-of-turn detection that knows a pause from a finished thought
- does things — web search (Tavily), Gmail, Calendar, Notion — every action that sends or changes something sits behind a permission gate
- remembers — searches every past conversation, saves and corrects her own facts about you, knows when you last talked and what was going on
- stays herself — six personalities, each a text file you can edit, tuned against a scripted conversation suite (short spoken replies, honest pushback instead of flattery, no guilt-trip goodbyes)
- Nemotron 3 Super is her mind: every conversation turn and every tool call (web search, memory, email). First word in about a second.
- Nemotron 3 Nano does her background thinking — reviewing what to remember and summarizing past conversations — at a fifth of Super's price, through the same Token Factory key.
- Reasoning switched off on the voice path (
enable_thinking: false). Measured on Token Factory: a reasoning pass put Nano's first word at 8.4 s instead of 0.8 s, and Nemotron 3.5 Lightning spoke its own "thinking process" aloud — unacceptable for a voice companion. - Cold-start warm-up: a cold serverless endpoint took 12.6 s to the first token (vs ~1 s warm), so Yumii sends a one-token warm-up while she boots and when you start speaking after a quiet spell.
- Live model list: the dashboard lists the tool-capable models your
Token Factory key can use (Nemotron Super, Nano, 3.5 Lightning, 3.5
Super VL, Ultra), straight from the
/v1/modelsendpoint. - NVIDIA Nemotron Speech (streaming speech recognition) is also selectable as an experimental on-device ear, via sherpa-onnx.
Token Factory's OpenAI-compatible API is why adding it took one wiring entry: Yumii's model catalog is data-driven, so Nebius plugged into the existing tool-calling agent without new code paths.
| Role | Options |
|---|---|
| Mind | NVIDIA Nemotron via Nebius Token Factory (default) · Groq · Gemini · OpenRouter · DeepSeek · xAI · Together · Mistral · OpenCode · Ollama (local or cloud) · OpenAI · Anthropic |
| Ears | Local Whisper (offline) · Groq Whisper (fast) · Vosk (offline streaming) · NVIDIA Nemotron Speech · Moonshine (both experimental) |
| Voice | Kokoro (local, free) · ElevenLabs · CAMB.ai |
| Web search | Tavily (with a key) · DuckDuckGo (free fallback) |
Everything switches from the in-app dashboard.
Private by architecture — no account, no telemetry, no server of ours. Your memories stay on your machine; conversation text goes only to the model provider you choose — or nowhere at all, with a local model.
- Proactiveness, done properly — reminders you ask for first, gentle offers instead of guesses, and silence as the default; researched and designed, not yet built
- Seeing your screen — she looks at what you're looking at and helps
- macOS & Linux shells · more voices and languages
The easiest contribution in open source: a personality is one text
file in src/yumii/assets/prompts/ —
copy one, rewrite the character, open a PR. Adding a speech or voice
provider is one self-registering module in src/yumii/audio/providers/
or src/yumii/tts/providers/ — see CONTRIBUTING.md.
MIT — see LICENSE. She's yours.