Jev in Practice: Browser Use and Agent Workflows

YoussefHosni95 explains that Jev is trained with RLCD, runs about 200x faster and 400x cheaper than most SOTA models, fills schemas with probabilities in parallel instead of generating tokens, and is still available by request only.

A user posts an exaggerated compliment about using Jev, with no concrete usage details beyond a personal impression.

A user asks whether Jev is handed to a browser_use agent as a tool so the agent can predict element refs and pick commands like click or press.

An open-source browser agent where Jev makes every decision and Vercel's agent-browser acts passed all 30 runs of 10 tasks, with a 3.1 s median and $0.0009 per task versus Claude Code's 9.4 s and $0.068.

A user says Jev simplifies routing and orchestration into questions and answers, combining a decision tree, classifier and plain-language understanding in one system.

After two stealth years, Diogo Almeida's Jev returns typed decisions with calibrated probabilities instead of generating text, running far faster and cheaper than an LLM, with a $40 million seed led by DCVC and no independent evaluations at launch.

A joking summary of Jev: models were scaled to trillions of parameters and fed most of the internet, all so the model could return a single enum.

A developer replicated Jev using DiffusionGemma and a lightweight Astra harness, landing within 2% on published benchmarks with a 0% hallucination rate and about $0.15 per million input tokens, with the code open-sourced.

Steve Krouse shared an online demo of Jev operating a browser via Kernel, highlighting its speed and inviting people to try it.

The author built a 254-token vocabulary from a trimmed GPT-2 set so Jev outputs one token at a time, with limited results and poor coherence.

After getting access, the author describes Jev as closer to a classifier than a chat model: write a question, send it, get a probability, and let code decide, with plans to revive an old project.

A researcher behind ChatGPT launched Jev, which returns choices, scores and probabilities instead of text, and the author lists eight uses including cheap first-pass judging, retry decisions, retrieval scoring and lead ranking.