Open JEV — zhihz

The second thing, and the project says so up front. The backend is a frozen Qwen3-4B-Instruct-2507, not a newly trained Open JEV foundation model, and there is no demonstrated performance lead over Jev or competing systems. It is an independent project inspired by Jev containing none of Jev's code, weights, or RLCD implementation, and the README states the name does not imply an official open-source release.

You define the decisions at request time, so no intent registration or per-question training is needed to submit something new, but reliable performance on arbitrary tasks is not guaranteed. Input is text only: no raw audio, images, PDFs, or live device feeds, which need an upstream transcription or extraction step. It returns decisions and does not send messages, operate devices, or execute tools itself.

Not established. Scores come from model logits rather than from the model writing numbers, and the README states that logit-derived scores are not automatically calibrated or better than other scoring methods. It also explicitly declines to claim higher accuracy than the same underlying LLM, universal latency or cost savings, calibrated certainty, or faster shared-context inference. Validate on your own data before setting thresholds.

Only on a setup close to the tested one. The quick start is validated on an Apple M3 with 16 GB unified memory and Python 3.14, and other platforms and dependency combinations are explicitly not validated. Budget disk for two environments: roughly 8 GB for the pinned source download and roughly 4 GB for the separate MLX 8-bit copy. Questions also run sequentially and re-encode the context.

Each question takes 2 to 8 candidates, and a question's full prompt can be up to 4,096 tokens. The web UI allows 16 questions; the API allows 100. Questions currently run sequentially and re-encode the context each time, so shared-context computation is planned rather than delivered, which means a many-question request will not behave like one parallel call.

Choice and Binary only. It returns a choice and relative probabilities over exactly the candidates you supply, inserting no extra answer or abstention option. It also returns stable candidate IDs alongside the scores, so an application can map a selected ID to a workflow branch without interpreting generated text. Input is text only: audio, images and PDFs need an upstream extraction step.

The project code is Apache-2.0, but the third-party models and datasets it uses keep their own licences, and no TypeSafe or Jev branding rights are granted. The README also says the bundle does not support full evaluation reproduction, so you can run the demo without necessarily recreating the authors' published numbers end to end.