Jev in Practice: Browser Use and Agent Workflows

Questions inside a single Jev request are evaluated in parallel, so packing every judgment you need at one moment into one request, one judgment per field, costs barely more than asking once.

The Japanese virtual character's official site uses Jev to judge comment type, mood, gesture and malice in parallel, picking reply text only from candidates so hallucinations and tone breaks are ruled out by design.

Deel compared Jev with frontier LLMs across several use cases, finding quality at parity or better on 6 of 8 checks, repeat-question accuracy rising from 70% to 97%, and costs up to 59x lower.

NeoHorse-Jev-4B is open-sourced with Choice, Noul and Score decision primitives, prefill-only inference, and a score of 77.70 across six text decision benchmark groups.

Typesafe says Jev outperformed frontier LLMs at a fraction of the price, with repeat-question matching rising from 70% to 97%, expense categorization from 50% to 86%, and fewer false alarms on escalations.

Kev is an open-source family of small decision models reverse-engineered from Jev's published architecture and built on Qwen, answering yes/no, multi-choice and rating in one call, with sizes from 0.8B to 27B that run locally for free.

Wang Yunhe's team at Jiyuan Ludong released NeoHorse-Jev-4B, a 4B decision model that scores candidates directly instead of generating text token by token, used for tool selection and routing, with a 74.3ms average decision time at 32 concurrency.

The author open-sourced his agent harness as the Keel repository, which integrates both Laya and Jev and can be handed straight to an agent.

An open-source JEV preview trading bot for Solana that routes spot orders through Jupiter and DFlow and trades Imperial Perps at 1x, running in dry-run only and requiring your own TypeSafe key.

An 18-year veteran developer released Bowser, a browser that builds itself, where users write their own mods generated with Jev and Astra, limited to the first 100 users.

A hook placed in front of an existing agent uses Grok 4.7 for writing and Jev in four seats for routing, judging and gating, with plain code as a fallback, all in one loop.

betaworks and Typesafe held the first Jev Night in New York, with demos and credits given away, which the author calls a sign that agentic reality is here.