Jev in Practice: Browser Use and Agent Workflows

A Thai-language explainer says TypeSafe AI's Jev takes data plus preset options and returns program-readable results such as type, urgency and confidence for tasks like email triage, and the company claims it is 40–200x faster than LLMs.

fastbrowse is an open-source browser agent where Jev picks each action directly from the page while an LLM plans, and every claim cites an exact quote; it is described as faster and cheaper, though early and experimental.

A video shows Jev handling browser use tasks at $0.0055 versus $0.04 for PrismML Bonsai 2 27B, with the post claiming Jev performed better than Google Gemma 4.

He argues that constraining an LLM to generate one token from a fixed set of options and using logprobs as a confidence score amounts to hand-building a Jev.

Jev never writes paragraphs, returning typed decisions like Choice, Score and Noul with confidence levels; the author uses an email-routing agent to show how to call it and its pricing.

A tutorial walks through 10 steps from creating an API key to wiring up a Cursor agent so Jev picks a model by cost, balanced or intelligence tier, claiming about 80% savings in cost and time.

A Thai developer feeds the browser Tetris board state and candidate placements as structured data to Jev to pick a plan, then the code presses keys and re-reads the screen to verify the result.

DJev, an Apache-2.0 open version of the Jev decision model built on Google's DiffusionGemma 26B, accepts images as options, runs text decisions in under 100ms and image-inclusive ones in 600ms, and is priced at $35 per billion input tokens.

A daily news roundup reports that TypeSafe's Jev dropped its waitlist and opened to all users, while a local open-source model called Laya mimicking Jev's architecture has also emerged.

A developer explored using TypeSafe's Jev with SQLite, DuckDB and LadyBugDB for hybrid search and re-ranking, and wrote up the work in a long post.

The open-source decision model Laya has been ported to Apple's MLX, running about 7-13ms per decision on an M3 Max with under 1GB of memory for local routing, scoring and guardrails.

LangChain reposted a simple guide covering what makes Jev special, how to use it with agents, and a few model use cases.