Open models bring Jev-style decisions straight to your browser

Amidst intense hype around fast-reasoning AI, an open experiment brings Jev-style decision making straight to your browser.

An open-source project called SemIf lets you run local decision benchmarks without sending data to a server. The site downloads quantized models—like MiniCPM5 2B or Qwen3.5 4B—from Hugging Face and runs them on your GPU using wllama. It tests two methods on the same prompt: reading option probabilities directly from logits, and writing probabilities token-by-token as JSON text. The project picked up 535 points and 242 comments on Hacker News.

Why it matters: Proprietary Jev models reach an 88.3% TypeSafe benchmark score, but they require a hosted API. Local open models don't quite match that accuracy yet—Qwen3 0.6B hits 40.7% while Qwen3.5 4B reaches 84.5%. However, running them in the browser gives you zero-backend privacy and precise execution timing measured on your own hardware.

Try it / know this: Head to the site and run MiniCPM5 2B on desktop or Qwen3 0.6B on mobile. The initial model download takes a few minutes, but the weights stay saved in your browser cache.

Your browser is officially a zero-latency AI sandbox.