Benchmarks & research
ytbai retrained OpenBMB's small MiniCPM5-2B to answer Jev-style choice, yes/no and score questions; the result runs on your own machine behind a /v1/systemone-style endpoint.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Rather than generating text, it reads the answer letters straight from the model's internal state, and it reads a shared document once for all the questions asked about it. The maker says no open model of 2B or fewer scores higher on public JevBench, though in the same table hosted Jev is still well ahead.
It suits a small team that wants to sort short texts, such as support messages or reviews, into labels it chooses, on its own computer rather than through a paid service. Several questions about the same message can be asked at once. Its maker's scores come from public test sets, so try it on your own examples before relying on it.