Jev is a narrow JSON classifier, not a frontier model

A breakdown of Jev, a startup from ex-OpenAI staff: the model does not chat, it picks from options, scores against criteria and returns true/false. Its claimed lack of hallucinations comes from masking illegal tokens at decode time, already used in OpenAI Structured Outputs and Anthropic tool use.

DEV Community Open source

Jev is a narrow JSON classifier, not a frontier model

Jev acts as a sidecar for routing, moderation, scoring and validation around a larger LLM
Context window is 32k, and on multiple-choice picks it reportedly lags a direct frontier-model query
Reported M4 speedups of 5.6–7.0x over autoregressive generation
Two hours after launch, Qwen-2.5-1B-RLCD shipped doing the same class of work

#OpenAI#Anthropic#Jev#Qwen

Jev is a narrow JSON classifier, not a frontier model

Ex-Google safety chief warns AI could harm children more than social media

Experts debate risk of an AI takeover of the internet

Breeze TTS 2 tops ElevenLabs on TTS leaderboard but bans commercial use

DeepSeek V4.1-Flash matches Claude Opus 5 on coding at a fraction of the cost