XNews

TypeSafe AI Launches Jev: Fast Structured Decision Model

TypeSafe AI Launches Jev: Fast Structured Decision Model

TypeSafe AI, founded by ChatGPT co-inventor, launched Jev, a System One model that outputs structured decisions with calibrated probabilities, 20-200x faster than LLMs. Cognition open-sourced Kev, a Jev-compatible model, achieving 79.6% accuracy on out-of-distribution data.

Key insights

Jev Design: No Free Text, No Hallucination

Jev does not generate free text; outputs are predefined structures with type safety enforced architecturelly, making type errors mathematically impossible. Every output includes an epistemically honest probability. This avoids hallucation inherent in RLHF where models learn to pretend confidence. Jev uses RLCD (Reinforcement Learning for Calibrated Decisions) training.

Performance and Cost Numbers

Jev response time 70-500 ms vs 3-329 s for frontier LLMs. Input cost $0.042 per million token, output free. TypeSafe claims 20-200x faster and 40-400x cheaper. Speed comes from parallel sampling instead of sequential token generation. Evals may underse Jev because referece is GPT-6 and Fable 5.1; extreme multiples (193x, 444x) are best cases.

Use Cases: Intelligent If-Statements and Agent Guards

Jev is designed for embedding in code as a structured decision node: classification, routing, scoring, extraction, guardrails. Examples: DOOM game AI ($7/hour), Wikipedia racing, game rendering via Manus. It complements LLMs in a three-layer architecture: LLM for open-ended work, Jev for fuzzy semantic decisions, deterministic code for safety limits.

Open-Source Kev Model

Cognition released Kev, a Jev-like model based on Qwen3 + LoRA + pointer head in three sizes (0.6B/4B/8B). Kev-8B achives 79.6% accuracy on out-of-tribution data vs Jev's 85.7%. It runs on consumer hardware (4B on 32GB Mac in ~300 ms). Training cost is very low (8B in 83 minutes on one H100). Fully compatible with Jev API.

Jecons Paradox and Pricing Strategy

The name Jev refs William Stanley Jevons' paradox: efficiency gains increase total consumption. TypeSafe bets that lowering decision cost to near zero will drive massive usage growth. Output token free, company already profitable and plans to cut prices further. The 'Bitterest Lesson' blog argues that task choice matters more than data and compute.

Source posts