OpenJev: browser-only demo pits reading a model's logits against generating JSON

OpenJev is a client-side experiment that runs small open-weight language models entirely in the browser—no backend, no waitlist—to demonstrate two ways of extracting a choice distribution from a model faced with a multiple-choice decision. The first method reads the model’s raw logits for the option tokens you supplied and applies a softmax across only those options, producing conditional probabilities without ever decoding text. The second asks the same model to write out an equivalent distribution as JSON, generating it token by token. Both paths run sequentially on one loaded model so they don’t contend for the GPU, letting you watch the latency and quality gap for yourself.

The demo defaults to MiniCPM5 2B (about 1.56 GB), with a lighter Qwen3 0.6B option for phones and a heavier 4B model for machines with more memory. Weights are pinned, quantized GGUF builds pulled from Hugging Face through wllama, cached locally in the browser; the page stresses that inputs never leave the device and that every stage—load, warmup, prompt prep, direct execution, first generated token, and completion—is timed live with performance.now() rather than shown as canned numbers.

The practical point is the case for ‘semantic ifs’: using a model’s option probabilities as branching logic while skipping the cost and variance of text generation. The authors are careful to caveat the results—the direct softmax scores are conditional over displayed options only, not calibrated confidence, and browser-side quantization can shift both accuracy and speed—so the tool reads less as a product claim than as a hands-on way to measure the tradeoff on your own hardware.