Frontier-grade intelligence that fits where frontier models can't.
Win by cunning, not by size.
Today's AI race is an arms race of size: bigger models, bigger clusters, bigger bills. Metis takes the opposite bet — a small reasoning core that wins through wisdom (reasoning), counsel (retrieved knowledge), and craft (tools), running entirely on hardware you already own.
01 / Cortex
A small, strongly-distilled reasoning model (Qwen3 1.7B–4B) that lives entirely in RAM. The brain that reasons but doesn't memorize facts.
02 / Library
Facts live on disk as a swappable corpus, not baked into weights. Swap the library and the assistant's entire knowledge changes — without retraining.
03 / Hands
Calculator, clock, web search via SearXNG. Offloads exactly what small models are worst at: exact arithmetic, fresh data, and long code execution.
04 / Conductor
Generate → Verify → Search. Never emits an unverified claim. Abstains rather than hallucinate. The same architecture that turned 0/8 into 8/8.
Your data, your model, your machine. Nothing leaves the device. Hospitals, courts, and individuals get capable AI without shipping their secrets to a cloud.
Useful AI on a $5/month VPS, a laptop, or an offline edge box. Not a $40k GPU node. Frontier-useful assistance in reach of the 90% of the world priced out of cloud AI.
The "brain" is fixed and small. What it knows is a file you can read, version, swap, and trust. Update the world's knowledge without retraining anything.
No internet, no API key, no rate limit, no vendor that can deprecate you. It just runs. One binary you can scp to any machine.
qwen3:1.7b called bare vs the same model inside Metis, on a private corpus it never trained on.
| Metric | Bare model | Metis |
|---|---|---|
| Answerable facts correct | 0 / 8 | 8 / 8 |
| Fabrications on unanswerable (lower is better) | 4 / 4 | 0 / 4 |
| General (incl. exact math) | 1 / 2 | 2 / 2 |
The bare model invents facts and confidently fabricates answers that don't exist. The same model inside Metis answers grounded, cites, and refuses to guess. Zero datasets. Zero retraining.
Every entry so far ran on borrowed weights. Today a model trained from scratch on this MacBook — on our own code, for zero dollars — answered inside OpenCode. It's tiny, it's bad, and it changes what this project is.
We tried to push Metis's founding bet — that verifying is cheaper than generating — to its logical extreme, and measured it. The bet is real, but it splits in two: recognizing a true answer is free, rejecting a plausible false one is not, and the second half doesn't scale with model size.
If the architecture does the work, the model should be replaceable. We swapped the 1.7B generator for a 0.6B one, 62% less model, and measured what broke.