Simon Willison on Jev: a 'decision model' that returns floats instead of text
Simon Willison · rss · 2026-09-22
Simon Willison reviews Jev from TypeSafe AI — billed a "System One model," better named a "decision model": text in, but floating-point numbers out for classifications, yes/no questions, scores, and confidences.
- Pricing: input-only billing, free output, $0.042/million tokens — cheaper than GPT-5 Nano ($0.05/M).
- Three question types: "Noul" (Bernoulli) yes/no with 0-1 confidence; choice questions with probability distributions; score questions along custom ranges. One state takes many questions in parallel, with latency similar to one.
- Use cases: classification tasks — spam detection, labeling, ranking; Willison tried search reranking (BM25 top-100 rescored by Jev).
- Black-box concerns: even more opaque than LLMs, no justifications, bias risks (he worries about ranking job applicants; its "good city?" Bay Area experiment rated Cupertino top, East Palo Alto bottom). Evals matter more, though running thousands of prompts costs cents.
- Community hacks: jevchat, jev-leftpad, jev-2048.
- Open-weight clones: Kev (Qwen 3.5-based, 0.8B/4B/9B) and a JevBench benchmark — impressive activity within a week of launch.
More from Models
- Anthropic investigates elevated errors across Claude Mythos 5.1, Fable 5.1 and Opus 5 — ClaudeAI-mod-bot · 2026-09-22
- Dev Swaps Opus for Mimo-v2.6 in Cline on Client Projects: 'It's a Beast' — MicahBerkley · 2026-09-22
- Kev refactored onto Qwen3.5: open-source decision models now at 0.8B, 4B and 9B — alexcovo_eth · 2026-09-22
- Why OpenAI bets on math: it's the most verifiable domain for reinforcement learning — burny_tech · 2026-09-22
- Open-source decision model Laya ported to Core ML: 99.5% ops on ANE, 3.7ms per decision — alexcovo_eth · 2026-09-22
- Fireworks: routing 18 models per task hits 97.6% solve rate at $1.88 vs best single model's 74.1% at $6.52 — sophiamyang · 2026-09-22