receptron/laya is a package that runs a Jev-compatible System-1 decision model — one that produces choices and scores rather than generating text — in Node.js and TypeScript. The latest release recorded on GitHub is v0.2.0. The provided information contains no changelog for this version, so the specific changes cannot be confirmed.
The model takes a state object and a typed question and returns three forms of calibrated probability — choice, score, and noul — in a single forward pass. The ticket routing example shows a department selection probability, an urgency score, and a churn risk probability. It runs on ONNX Runtime with Node.js 20 or later and does not require a Python runtime. The README states that its output matches a Python reference implementation to four decimal places. So rather than a text generation model, it is a decision model that adds routing and classification to agent pipelines.
That said, the fp32 weights are about 1.7 GB and are not included in the repository. They are downloaded from Hugging Face on first use and cached locally, and the output parity check only runs in an environment where the weights are ready. The performance evidence is limited to a single environment measurement: a cold download and a 140 millisecond warm latency. CI checks cover several operating system and Node version combinations, but it remains difficult to judge real-world speed from the performance figures alone.