---
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.*