Particle.news

TypeSafe AI Unveils Jev, a Non‑Text 'System One' Model for Fast Structured Decisions

Jev returns Choice, Score and Noul primitives with confidence estimates to speed real‑time scoring and routing as the company supplies vendor benchmarks that remain unverified.

Overview

  • TypeSafe AI announced Jev as a purpose‑built model that does not generate natural language or code but outputs three structured primitives—Choice, Score and Noul—each with calibrated probability or confidence values.
  • The company says Jev targets high‑throughput, low‑latency tasks like risk scoring, moderation routing and model verification by returning compact decisions instead of autoregressive text.
  • TypeSafe billed a new training method called calibrated decision reinforcement learning (RLCD) intended to align the model’s reported confidence with actual accuracy for better calibrated probabilities.
  • TypeSafe published vendor benchmarks claiming large speed and cost advantages and released pricing that charges $0.042 per million input tokens while treating outputs as effectively free; those performance and cost figures have not been independently verified.
  • The company exited stealth with a $40 million seed led by DCVC, showed a real‑time demo of Jev playing Doom at roughly 10 calls per second, and positioned the model as a tool that could lower automation costs and enable real‑time pipelines for businesses.