Benchmarks & research
AgentJev-0.6B is a tiny model for coding agents that answers questions like 'is this task really finished?' with a probability for each option, without writing any text.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
malevrigns took Qwen3-0.6B, removed its ability to write text, and added a small scorer that rates whatever options you list, in any order. The weights now on Hugging Face are tuned for judging finished coding work; the README's broader typed-decision results come from an earlier version. Anything longer than 2,048 tokens is refused rather than trimmed.
Picture an automated coding helper that says it has fixed a bug. AgentJev can be the quick second opinion: shown the request, the change and the test results, it says how likely the job is really done. Its maker warns that this is not the same as the change being accepted, so compare it with your own past work first.