The omni model small enough to be yours.

The frontier is racing to build giants in datacenters.
moche runs the other way.

Every year models grow, and every year they drift further from the machines people actually own. moche is a bet on the opposite corner of the map: one small unified model — understanding and generation in a single interface — that fits where you already work.

Not a demo of a distant future. A thing you download, and it's yours.

One brain. Four organs.
Stitched with learned connectors.

A frozen multimodal backbone does the thinking. Generation is grafted on through latent connectors — the brain's inner state flows directly into each decoder, no lossy text bottleneck in between. Speech lives inside the brain itself, as codec tokens.

┌────────────────  moche · ~8B class · ships Q4  ────────────────┐
│                                                                  │
│   [ brain — 4B multimodal LLM, frozen ]                          │
│      │                                                           │
│      ├── speech codec tokens ─→ codec decoder     voice        │
│      ├── <gen_image> → queries → connector → DiT   hand         │
│      └── <gen_video> → connector → video DiT      motion       │
│                                                                  │
└──────────────────────────────────────────────────────────────────┘

Runs where you are.

Distributed as MLX bundles for Apple Silicon. CUDA builds follow.

Small isn't a port.
It's the design constraint.

Every stage of moche has one gate: survive Q4 quantization or don't ship. And the training data isn't scraped — it's farmed. An orchestrated generation pipeline produces candidates, vision-language critics score every one, and only the winners enter the training set.

Four stages. Each one ships.