laya_router: a model router with Laya matches GPT-5 nano's accuracy, 35x faster, free

Your model router runs on every request. Does it need to be a language model?</h2>

[](https://github.com/glukicov/laya_router/actions/workflows/ci.yml)
[](LICENSE)
[](.python-version)
[](https://github.com/astral-sh/uv)
[](https://github.com/astral-sh/ruff)
[](https://github.com/astral-sh/ty)
<br>
[](https://huggingface.co/convaiinnovations/laya)
[](https://platform.openai.com/docs/guides/structured-outputs)
[](https://fastapi.tiangolo.com)
[](https://kind.sigs.k8s.io)

**[Read the accompanying article on Medium](https://medium.com/@lukicov/smart-routing-with-an-open-source-system-1-model-laya-416fe771ce2e) · [Result](#result) · [Quickstart](#quickstart) · [The experiment](docs/EVAL.md) · [The data](data/README.md)**

</div>

**laya_router — a smart model router, with two brains.** Every request goes to the router first: it decides
whether a `small`, `medium` or `powerful` model should answer, then the work goes there. That decision is on
the critical path of every single request, so the router's own latency and bill are pure overhead.

Which makes routing a **System 1** job. In Daniel Kahneman's
[*Thinking, Fast and Slow*](https://en.wikipedia.org/wiki/Thinking,_Fast_and_Slow), System 1 is fast,
automatic and intuitive; System 2 is slow, deliberate and effortful. Deciding *which model should answer
this* is a reflex, not a deliberation — you want the snap judgement, and you want it before the real work
starts.

This repo puts two routers behind one endpoint and measures them on 180 labelled requests.
[**Laya**](https://huggingface.co/convaiinnovations/laya) is a 421M non-autoregressive decision model that
describes itself as a *System 1 decision engine* — its API method is literally `system_one()`. It answers
all three routing questions in **one forward pass** as prob