TypeSafe launches Jev model for structured programmatic decisions
TypeSafe, founded by OpenAI veteran and ChatGPT co-inventor Diogo Almeida, has introduced Jev after two years of stealth development. The company describes it as a specialised System One Model for making structured probabilistic decisions inside production software, rather than generating text. It accepts unstructured state and returns typed values through a single parallel query.
Jev is trained with Reinforcement Learning for Calibrated Decisions (RLCD), an approach intended to align confidence scores with the accuracy of execution decisions. Its hardware-aware sampler evaluates structured options in parallel and supports selections with up to 255 discrete choices. According to TypeSafe, predefined output schemas remove the need for parsing pipelines and additional guardrails used with conversational language models.
Internal tests reported response times of 70–500 milliseconds, compared with 3–329 seconds across conversational frontier deployments. Workflow evaluations recorded speeds up to 193.6 times faster. Input processing costs $0.042 per million tokens, while output tokens are not metered. Demonstrations included a Doom bot running at 10 queries per second for about $7 per hour and Wikiracing tests. TypeSafe has opened early developer access and begun onboarding teams from its waitlist.
Written by Sakhanda Wire AI from reporting by AI News. Facts, quotes and full context are in the original article below.