openjev presents a browser-based experiment exploring whether a decision model can run locally without backend dependencies. The project allows users to select model sizes, including Mini CPM 5 2B as default or Qwen3 0.6B for smaller devices, and measure performance differences between direct probability readouts and token-by-token generation. Users load quantized weights from Hugging Face, which remain cached locally, with inputs staying entirely on the page.

The experiment compares two methods: direct readout, which normalizes softmax across supplied options without decoding, and generation, which asks the model to write option probabilities as JSON text token by token. Both run sequentially on the same loaded model to avoid GPU contention. Weights originate from Hugging Face and persist in browser cache.

First model loading may take several minutes based on selection, network, and hardware. The demo uses pinned GGUF builds through wllama, noting quantization affects both quality and speed. This local experiment offers immediate access without waitlists, enabling users to benchmark decision model performance directly in their browser environment.