What it is
Serves a contrastive language model that scores actions against states for typed decisions, ranking, and verifier use with much lower latency than Jev. Score candidate actions against a state, then use the same primitive for typed questions, best-of-N ranking, tool routing, and verifier selection through a local API or in-process engine.
Key results
- Up to 9× lower latency than Jev on computer-use, gaming, tool-calling
- Verifier fine-tuning hits 87.6% Terminal-Bench 2.1 and 81.6% DeepSWE
- Repeated-state serving drops p50 from 1.7-2.0 ms to 0.6-0.7 ms on RTX 4090
Shows that contrastive state-action scoring can replace generative selection for many agent decisions while being much faster.