AI-Wiki

TypeSafe AI

3 min read

TypeSafe AI

Confidence 0.70 · 3 sources · last confirmed 2026-09-22

TypeSafe AI is a startup that in September 2026 released Jev, which it calls a System One model: “a class of AI models built to make fast, structured decisions that software can use directly. A System One model evaluates a state and returns typed answers and probabilities.” The name borrows Kahneman’s System 1 / System 2 distinction. Jev is cast as fast intuition; generative LLMs are the slow, step-by-step reasoning.

The wiki knows TypeSafe only through LangChain. All three sources are LangChain publications about the integration. No TypeSafe primary source (its launch post, docs or model card) has been ingested. The alias Jev is recorded here because the wiki has no separate product page for it.

What is claimed about Jev

  • Interface. A state (text, structured data or chat messages) plus typed questions of three kinds: Choice (pick one option), Score (rate on ordered levels) and Noul (yes/no probability). All questions in one request are evaluated in parallel.
  • Training. “Reinforcement learning for calibrated decisions (RLCD).” Size, architecture and training data are not disclosed.
  • Speed and cost. “Up to 200x faster inference and 400x lower cost than comparable LLMs on classification tasks”, or as a range, “20 to 200 times faster and 40 to 400 times cheaper”. These are TypeSafe’s figures, repeated by LangChain without a method.

What has been measured

Only one measurement exists in the wiki: LangChain’s Jev-as-a-Judge experiment. On five fixed weather-agent traces, Jev agreed with a human rater on all 500 binary judgements and had by far the lowest score variance. It cost $0.00035 per call against $0.00039 for GPT-5.6 Luna and $0.02811 for Claude Sonnet 4.6, at 0.44 s latency against 2.16–2.83 s. That is 1.1× to 80× cheaper and ~5–6× faster on this setup. See that page’s scope section before citing it.

Role in the wiki

Concepts touched: small-language-models (specialised models inside heterogeneous agent systems), agent-development-lifecycle (judges), agent-oversight-and-delegation (risk classifiers).

Open questions

  • Independent evidence. Every source is from a distribution partner. A TypeSafe primary source, or a third-party benchmark, would be the next thing to ingest.
  • Is it a small model? The wiki files Jev under the specialised-model argument, not the small-model one, because its size is unknown.

Graph View

Backlinks