LitJev

What LitJev does

LitJev is an independent research reproduction of the Jev decision layer on Hugging Face Qwen checkpoints. You load a model, define typed questions, and read scores from the output head — no training loop and no generated answer text for the decision itself. The HTTP surface mirrors the public System One schema (POST /v1/systemone with model / state / questions), so client code written against Jev can often switch base URLs for experimentation. Default testing targets large Qwen3.x checkpoints on GPU; vision checkpoints are required for screenshot decisions. Probabilities are not calibrated by default, and the project is explicitly not affiliated with TypeSafe. Use LitJev when you need a self-hosted, Jev-shaped endpoint for research, demos, or offline evaluation — not as a drop-in guarantee of identical confidence or latency to hosted Jev. LitJev is aimed at researchers and platform teams who want a self-hosted decision endpoint that looks like System One on the wire while remaining honest that internals, confidence, and latency will not match hosted Jev.

Where it fits in the Jev stack

• Self-hosted System One–shaped APIs for internal experiments
• Comparing local Qwen decision quality against hosted Jev clients
• Browser-facing demos that need Choice / Score / Noul without a TypeSafe key

Browse more in Benchmarks & research, or return to the Jev AI tools directory.