TypeSafe AI Releases Jev, a System One Model for Structured Outputs

Jev: Structured Decisions as a Model Primitive

Jev is the first public model in TypeSafe AI's System One Model class, targeting fast, structured decisions that software calls directly rather than generating text for downstream parsing.

• Input/output: takes unstructured program state, returns type-safe structured values with calibrated confidence scores and probabilities — usable as a fuzzy decision rule for classification, routing, scoring, extraction, or branching.
• Training: Reinforcement Learning for Calibrated Decisions (RLCD).
• Inference: parallel sampler emits all outputs in a single pass instead of token-by-token.
• Performance: TypeSafe reports 70–500ms end-to-end latency on System One tasks, versus 3–329 seconds for comparable frontier models.
• Pricing: $0.042 per million input tokens, no output token cost.
• Schema matching: guaranteed by construction rather than validated post-generation.

Jev is available today in early access.

Link utils

• Fake Jev demos are taking over the internet
• Will TypeSafe’s Jev change how we build AI applications?
• Building a Harness with Jev
• The NEW Blazing Fast AI Model: Jev
• A deep dive into Jev, TypeSafe's System One model