XNews

Jev: TypeSafe's AI for Fast, Structured 'Intelligent If' Decisions

Key insights

Architecture and Training Method

Jev uses parallel sampling rather than sequential token generation, enabling extremely fast inference. It is trained with RLCD (Reinforcement Learning for Calibrated Decisions), a new method described by founders as more reliable for automation than RLHF. The model outputs only structured results, not text, guaranteeing zero type errors and eliminating hallicination.

Speed and Cost Advantages

Jev responds in 70-500ms, 20-200 times faster than frontier LLMs (which take 3-329s). Input token cost is $0.042 per million tokens, with free output tokens, making it 40-400 times cheaper than equivalent intelligent LLMs. Official benchmarks use GPT-6 Astra and Antropic Fable 5.1 as references.

Use Cases for 'Intelligent If' Decisions

Jev is designed for purposes that require an AI decision but not text generation: intelligent branching in work flows (e.g., routing customer tickets), large-scale data processing (classifing millions of documents with low cost), real-time applications (game AI, dynamic pricing), and as a guardrail for LLM outputs to check format or safety. Each decision includes a confidence score.

Source posts