Jev, a judgment-only AI model from an ex-OpenAI researcher, runs 200x faster and 400x cheaper

量子位 · wechat · 2026-09-20

TypeSafe AI, founded by ex-OpenAI researcher Diogo Almeida (RLHF, InstructGPT), unveiled Jev, a "System 1" model that skips text generation entirely and only outputs decisions, scores, and calibrated probabilities.

Three primitives: Choice, Score, and Noul (probability that a judgment holds) — effectively a probabilistic "semantic logic gate"; independent questions run in parallel.
Speed/cost: 70–500ms end-to-end responses, 20–200x faster and 40–400x cheaper than frontier models; input $0.042 per million tokens, output free forever.
RLCD: a new training method (Reinforcement Learning for Calibrated Decisions) aligning output probabilities with real-world frequencies.
"Zero hallucination" caveat: it never invents invalid options, but can still pick the factually wrong one.
Developer use cases: safety classification (5–18x faster than GPT-5.6 Luna), email triage vs Gemini (slightly less accurate but 10–20x cheaper), low-cost LLM verifiers, jailbreak monitoring, and model routing. A chess test showed Jev at 0.3s/$0.0001 per move still lost to GLM 5.3 — hybrid routing is the takeaway.
Open-source Nimble: LoRA adapter for Qwen3.5-9B (Apache 2.0), trained with a spot-the-difference data recipe; agreement with reference labels 90.12% vs base model's 66.36% (Jev: 93.21%).

Original post →

Jev, a judgment-only AI model from an ex-OpenAI researcher, runs 200x faster and 400x cheaper

More from coding & agent