Laya and Jev: Open-Source Decision Models on GitHub
Laya and Jev are two names for the same idea: a small model that answers a typed question and returns a calibrated probability, instead of writing prose. TypeSafe ships Jev as a hosted API; Laya is the open-source implementation of the pattern, and the projects below are what developers built around both.
NandhaKishorM/laya
5.2k★ · Apache-2.0 · v0.3.4 (2026-09-20)
Laya — an open-source, non-autoregressive “System One” decision model family. 421M parameters, a single forward pass (~33 ms), 100+ languages, typed questions and calibrated probabilities instead of generated text, so there is no free-form output to hallucinate. Ships with a PyPI package (pip install laya), a router mode that picks the right preset for the incoming script, and model cards on Hugging Face under convaiinnovations.
browser-use/jev-ultrafast
12.8k★ · MIT · created 2026-09-16
A browser agent that leans on TypeSafe Jev’s speculative fan-out: instead of one branch at a time, it fires several candidate actions in parallel and keeps the one the decision model scores highest. It is the fastest-growing repository in the Laya / Jev orbit, which is why it shows up on the trending boards.
wotai-dev/typesafe-jev-tools
Shell · created 2026-09-18
A Claude Code PreToolUse hook that asks whether a step needs a language model at all before spending one, plus a measured 149-row comparison of Jev against Claude Haiku 4.5 — the sort of independent measurement the hosted-model launch posts did not include.
Why Laya and Jev keep showing up together
The pitch is the same in both cases: routing, triage, moderation and guardrails are classification decisions, not writing tasks. A decision model answers them in one forward pass, so it is cheaper and far faster than a chat model, and because the answer space is enumerated, there is nothing to hallucinate. TypeSafe published the pattern and a hosted API; Laya published an open 421M-parameter implementation with its training and evaluation setup.
The open question is benchmark honesty. The hosted launch numbers and the open model numbers are measured differently, and the most useful repository in this list is the one that measured both side by side. If you are evaluating a decision model for production, read that comparison before the marketing pages.