OpenJev is a browser demo that measures reading option logits directly against making the model write JSON probabilities

OpenJev (via Hacker News, 179 points)·medium signalXBlueskyLinkedInCopy link

OpenJev runs a decision model entirely in the browser on WebGPU with no backend, loading Qwen3 0.6B (639 MB), MiniCPM5 2B (1.56 GB) or Qwen3.5 4B (3.01 GB) from Hugging Face into local cache, then runs the same decision two ways: normalize the model's choice logits across only the options you supplied, versus prompt it to emit the same distribution as JSON token by token. The published TypeSafe numbers on a 102-row subset are 40.7% for Qwen3 0.6B, 63.7% for MiniCPM5 2B and 84.5% for Qwen3.5 4B, against 88.3% for the hosted Jev result. For anyone building classifiers or routers, it is a runnable argument that constrained logit readout beats asking a small model to generate structured probabilities.