XNews

TypeSafe AI launches Jev System One model for fast decisions

TypeSafe AI launches Jev System One model for fast decisions

TypeSafe AI launched Jev, a System One model for fast decisions. It outputs structured results in 70-500ms with calibrated confidence, 20-200x faster than frontier LLMs. Input costs $0.042/M tokens, output free. The model cannot generate free text, eliminating hallucination.

Key insights

Jev uses parallel sampling to achieve millisecond decision speed

Unlike autoregressive LLMs, Jev uses parallel sampling to compute all outputs at once, enabling 70-500ms response times. It is 20-200x faster than frontier models like GPT-6 Astra and Claude Fable 5.1. The cost advantage is 40-400x cheaper.

Jev eliminates hallucination by never generating free text

Jev outputs only predefined structured types with associated confidence scores. Type errors are mathematically impossible due to architectural constraints. This design is a response to RLHF's tendency to force model confidence, which leads to hallucination. Jev can only express calibrated probabilities.

New RLCD training focuses on calibrated decisions

Jev is trained using Reinforcement Learning for Calibrated Decisions (RLCD), a method developed by TypeSafe CEO Diogo Almeida. Unlike RLHF, RLCD does not optimize for human approval but for accurate, well-calibrated decisions. This directly addresses the reward model asymmetry that causes overconfidence in LLMs.

Jev designed for classification, routing, scoring, and extraction tasks

The model is meant to replace brittle rule engines or expensive LLMs for tasks like intent classification, document tagging, real-time recommendations, and guardrail validation. It can be embedded in code pipelines as a judgment function. Demonstrations include controlling a DOOM bot at 10 queries/second ($7/hour) and Wikipedia racing.

Free output tokens and low input cost aim to drive massive usage

Jev charges $0.042 per million input tokens and free output. The pricing reflects the Jevons Paradox: as decision costs drop, usage multiplies. TypeSafe claims it is already profitable and plans further price cuts. This contrasts with token-metered LLMs.

Jev not a coding agent; CEO critiques current agent design

TypeSafe CEO argues that many coding agent features are workarounds for KV cache limitations. Jev is not a coding agent but can be used as a component. He proposes a "TypeSafe-centric" agent design with explicit state management and dynamic context reconstruction, which could reduce costs and improve routing.

Jev in early access, available via waitlist

The model is released in early access. Developers can join a waitlist. No full public availability yet.

Source posts