TypeSafe AI Introduces System One Models and Jev for Fast, Structured Decisions

Summary

TypeSafe AI has announced System One Models, a model category designed to make fast, structured decisions that software can use directly, and has opened early access to its first model, Jev. Founder Diogo Almeida says the company spent two years in stealth developing a new architecture, a parallel sampler, and a training method called Reinforcement Learning for Calibrated Decisions (RLCD). Jev does not generate strings; it returns predefined, type-safe structured values together with probabilities and confidence scores. TypeSafe claims that Jev reaches similar intelligence levels to existing LLMs on System One tasks while being two orders of magnitude faster and more efficient, with reported end-to-end latency of 70 to 500 milliseconds and input pricing of $0.042 per million tokens; output is described as free. The company positions the model for classification, routing, scoring, extraction, branching, real-time applications, workflow automation, and verification of other models. Its technical comparison contrasts sequential token sampling and human-preference or verifiable-reward training in LLMs with parallel sampling and RLCD, which optimizes calibrated decisions. TypeSafe says Jev cannot produce type errors because its schema-constrained outputs are mathematically guaranteed to match the declared structure, although its claim that the model avoids hallucination is presented with more limited empirical support. In workflow evaluations, every model used the same code-defined workflow, while the average predictions of external frontier models served as reference probabilities. TypeSafe reports that Jev occupied the Pareto frontier across nearly two orders of magnitude, but acknowledges possible bias because the workflows were created by its capabilities team and the reference models and wrapper may affect the comparison. Demonstrations included a reactive Doom bot and Wikiracing, with Jev supporting choices of up to 255 options and using a two-stage process for higher-cardinality decisions. The company is bringing developers off its waitlist gradually and says it plans to expand the system while collecting feedback on where Jev succeeds or falls short.