OpenJev

OpenJev is a browser-only experiment that runs a Jev-style decision model locally and compares two ways of extracting a model’s choice distribution: a direct readout of logits over supplied options, and a token-by-token generation where the model writes option probabilities as JSON. It lets users load models into the browser (weights pulled from Hugging Face and cached locally), run both methods on the same GPU sequentially, and measure real wall-clock timings for setup, warmup, first token and full generation. The demo emphasizes privacy and immediacy - inputs never leave the page and there’s no backend or waitlist - so anyone can load a model and see the timing and output differences on their own hardware.

The lab offers three model tiers (Qwen3 0.6B at 639 MB, MiniCPM5 2B at 1.56 GB, and Qwen3.5 4B at 3.01 GB) and reports per-model performance metrics (for example MiniCPM5 shows ~68.6% authored, 69.3% perturbed, 63.7% typesafe; Qwen3.5 ~81.3/76.6/84.5) against a published Jev typesafe reference of 88.3%. It notes important caveats: direct scores are softmaxed only over displayed options and aren’t calibrated confidence, quantized browser builds (GGUF via wllama) can affect speed and quality, smaller-phone models trade accuracy for size, and none are claimed to fully match published Jev. The interface shows measured wall-time ratios and lets users inspect token streams during generation.