Laya on Bun

A monorepo that runs Laya locally on Bun: an HTTP inference server with native ONNX inference and Hugging Face tokenizers, plus a bilingual English/Persian playground with 16 interactive demos. Python is needed once for the model export; serving is all Bun.

The server implements TypeSafe's decision API, so you point the official SDK at your own baseURL — for example new TypeSafeClient({ baseURL: 'http://127.0.0.1:3000' }) — and call client.systemOne with a state and typed questions.

choice returns the selected label and probabilities for all options, score returns an expected score on a zero-based ordered scale, and noul returns the probability that a statement is true. Laya does not generate text.

Bun 1.4.2 (the version pinned in CI), Python 3.12 with uv, and several GB of disk and RAM: the English source weights are about 820 MB and the float32 ONNX model about 1.6 GB, before export dependencies.

Install dependencies with bun install --frozen-lockfile, run the download and export scripts once, then keep two terminals: bun start for the inference API on port 3000 and bun run playground for the playground on port 3001. The API is ready once the model has loaded.

bun run smoke starts its own temporary HTTP server, checks JavaScript/Python token parity and compares real SDK responses against the exported reference fixtures. It does not need a separately running server.

The repository declares no license file, so check before production use. If you already have a complete exported model directory, set LAYA_MODEL_DIR to its path and skip the download and export steps; serving never needs Python.