Kev: open-source Jev-like decision models from 0.8B to 27B run locally for free
alexcovo_eth · x · 2026-09-26
The community released Kev, an open-source family of small decision models built on Qwen3.5/3.8 and reverse-engineered from Jev's published architecture, aimed at offloading agents' classification-heavy calls from expensive large models.
- Drop-in compatible with typesafe's SDK; answers yes/no, multi-choice, and rating in one call
- Sizes range from 0.8B (runs on a laptop) to 27B (data-center GPU), with a zero-install browser demo
- The companion JEV 101 guide notes a $0.03-per-call model spends most of its time on judgment tasks that don't need generation — local small models can do them for free
More from coding & agent
- New AI Coder Meme: Live Long Enough and You End Up Building a Harness — justalexoki · 2026-09-26
- ACuRL: zero-human-data continual learning for computer-use agents lands at NeurIPS — ysu_nlp · 2026-09-26
- pr-lens fork re-creates the missing open-source canvas for animated PR walkthroughs — jasonkneen · 2026-09-26
- Andrew Ng: AI engineering tactics must calibrate to project stage, from evals to architecture — DeepLearningAI · 2026-09-26
- Running multiple Claude/Codex sessions on NPM? Switch to pnpm and save tens of GB — altryne · 2026-09-26
- One Skill for World-Class Agent Research: Multi-Source Verification and Skeptic Voting — EXM7777 · 2026-09-26