The field of decision-oriented AI models is seeing rapid activity following the announcement of Jev by TypeSafe AI on September 15, 2026. Jev, described as a "System One Model," is designed for fast, structured decision-making rather than conversational text generation. By utilizing a parallel sampler and a training method called Reinforcement Learning for Calibrated Decisions (RLCD), Jev outputs typed, probabilistic decisions in a single forward pass. This approach allows it to achieve response times between 70ms and 500ms, and the company claims it can be up to 193.6x faster and 444.6x cheaper than frontier LLMs for specific tasks.
Model ReleasesJevTypeSafe AId1Liquid AI
Decision-making models gain momentum with TypeSafe AI's Jev and Liquid AI's d1
The release of Jev has catalyzed the development of similar architectures. Jared Palmer released "Kev," a family of small decision models built on Qwen3.5 that are compatible with TypeSafe's System One API. More recently, Liquid AI announced "d1," a decision model that aims to build upon the momentum established by Jev. According to Liquid AI, d1 outperformed Jev on the Hugging Face Decision Index, demonstrating superior performance in multilingual evaluations, greater robustness against prompt injection, and more effective handling of longer inputs.
While traditional LLMs are optimized for generating sequential strings, these new models focus on providing "epistemically honest" probabilities and confidence scores for structured values. This focus on type-safety and high-speed execution aims to make AI more suitable for direct integration into software workflows, such as customer support routing, real-time decision engines, and automated data classification.
Sources
- 「Jev」とは “文章を書かないAI”がなぜ話題に？ 元OpenAI研究者が開発、「“スマートなif文”と考えてみて」 （ITmedia AI+, 2026-09-20）
- TypeSafe AI Blog
6 more sourcesHide sources
- Kev: Tiny Jev-like family of decision models built on top of Qwen3.5 （Hacker News Frontpage, 2026-09-21）
- Kev-4B
- Jev-Leftpad （Hacker News Frontpage, 2026-09-21）
- Documentation
- ハルシネーション排除、最大200倍高速――元OpenAI研究者の新AI「Jev」は何が違うのか （ITmedia AI+, 2026-09-25）
- ベンチマークでJev超えな意思決定モデル「d1」が登場 （GIGAZINE, 2026-09-30）