Jev was built for agents, here's how we're using it in computer use instead

Jev is a new class of model built as a “foundation model for classification”: it consumes unstructured state and returns typed, probabilistic decisions (Choice, Score, Noul) with calibrated confidences. Choice picks one option from up to 255 with per-option probabilities, Score ranks content against descriptive levels, and Noul returns a yes-probability. Trained with reinforcement learning from calibrated decisions (RLCD) rather than RLHF, Jev is optimized for trustworthy probability estimates and parallel generation, yielding much lower latency and token costs than autoregressive LLMs. It offers a 64k-token context window (state + longest question limited to 32k), internal benchmarks claiming superior accuracy-per-cost and accuracy-per-speed versus major LLM vendors, and a pricing profile that makes high-volume classification affordable.

Jev is not intended to replace generative agents but to be embedded as a deterministic decision primitive inside software: customer-support routing, invoice processing, security triage, agent monitoring, and other automation tasks. One practical integration class uses Jev to pick UI selectors and actions inside the Stagehand framework: Jev classifies the requested action, ranks candidate targets, and accepts a candidate above a 0.7 threshold while falling back to an LLM if needed. Early testing cut median Act latency from 1.97s to 0.46s (≈4.3× faster), demonstrating how Jev can make AI-powered systems faster, cheaper, and more predictable without relying on full generative reasoning.