OpenJev
Points and comments are a snapshot, not live.
OpenJev runs a local decision model in the browser using GGUF-quantized LLMs.
OpenJev lets users load models like Qwen3 0.6B (639 MB), MiniCPM5 2B (1.56 GB), or Qwen3.5 4B (3.01 GB) to compare direct logit readout with token-by-token JSON generation for the same decision. All processing is local via WebGPU and wllama; inputs never leave the page. Model quality varies: on a public 102-row subset, MiniCPM5 2B scores 68.6% authored and 69.3% perturbed accuracy, while the published Jev scores 88.3% on TypeSafe.
What commenters are saying
Commenters note Jev is a specialized classifier model, not a general LLM, and that OpenJev's approach of restricting output tokens and using small models is not novel. Some question the naming, as it is not affiliated with Jev. One user replicated similar functionality in Rust/CUDA. Others criticize the LLM-generated site design as off-putting, though some find it clean. A user tested a real-world choice and got a different answer than expected, sparking debate on model suitability.