File size: 3,408 Bytes
29023e2 8a0fb8a e93ff1d 8a0fb8a cb4dac9 8a0fb8a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 | ---
license: apache-2.0
language:
- en
tags:
- on-device-ai
- privacy-first
- mlx
- asr
- speech-recognition
- local-llm
---
# Yooz Labs
**Sovereign Intelligence. Built for the skeptical.**
Privacy-first AI that runs entirely on your devices. No cloud, no tracking, no compromises.
---
## What we're building
We're the privacy infrastructure for the AI decade. Every Yooz product is designed for the 70% of consumers who don't trust cloud AI but lack consumer-grade alternatives.
- **Yooz Engine** — unified local AI service for macOS (STT, LLM, grammar, VAD, TTS).
- **Yooz Whisper** — voice keyboard for macOS.
- **Yooz Notes** — note-taking with private AI memory.
- **Remi** — Claude Code's distant friend. Secure peer-to-peer remote sessions for Claude Code (and soon Codex), with an iPhone app and local auto-approve.
- **Yooz Vault** — privacy hardware (home server).
- **Universal AI Platform Layer** — one API across Apple Core ML, Android ML Kit, Windows DirectML.
## What lives on this Hugging Face org
The **model weights** — Apache 2.0, fully open source. The Yooz product code is source-available on GitHub under PolyForm Shield, but the weights stay open so the research community can build on them, audit them, and remix them.
### Model categories
| Category | What it is |
|---|---|
| **ASR** | Speech-to-text checkpoints (Qwen3-ASR Swift port, Parakeet derivatives) |
| **LLMs (Touchup)** | Fine-tuned small LLMs that fix/clean speech-to-text transcripts |
| **Distillations** | Small students distilled from larger teachers for on-device inference |
| **Adapters** | LoRA / DoRA adapters published alongside their fused checkpoints |
All checkpoints document their **lineage** (base model + Hugging Face link), **eval numbers** (real benchmarks, not vibes), and **Swift / Python usage snippets**.
## Why open weights?
The competitive moat in privacy-first AI lives in the **product**, not the weights:
- Multi-device orchestration (phone → PC → Vault) over WireGuard mesh.
- Universal platform abstraction across Apple, Android, Windows AI APIs.
- Private AI memory: encrypted, local, with permissioned cross-app context.
- Beautiful, consumer-grade UX.
The weights themselves should be open so the research community can audit privacy claims, reproduce evals, and build on top. Releases follow the standard "ship the artifact + the recipe to reproduce it."
## Provenance
We never train on user data without opt-in. All training corpora and synthetic data sources are documented in the model card for each checkpoint. Where we fine-tune from a base model (Qwen, Gemma, etc.), the lineage is preserved and the upstream license is respected.
## Source-available, not closed
The product code lives at [github.com/yooz-labs](https://github.com/yooz-labs) under PolyForm Shield. You can read it, fork it, and build on it for non-competing use cases. We chose this path because we want to stay community-aligned without enabling AWS-style "managed Yooz" competing services.
## Get in touch
- **Engineering & research**: dev@yooz.info
- **Bugs and feature requests**: file on the relevant GitHub repo under [yooz-labs](https://github.com/yooz-labs)
- **Mailing list / news**: coming soon
---
*We're building the privacy infrastructure for the AI decade. Every decision prioritizes user sovereignty, data privacy, and beautiful simplicity.*
|