Contrastive-LM releases CLM-8B, an open model scoring actions up to 9x faster than Jev

The release targets the same interface as Jev, a proprietary System One model from TypeSafe AI that entered limited early access on 15 September 2026. CLM-8B exposes three question types: Noul (truth probability), Choice (option selection), and Score (rubric evaluation). In zero-shot tests, the research team reports CLM-8B runs up to 9 times faster than Jev, particularly in tasks with repeated actions like the T-Rex game. While it matches Jev on some tasks, it trails on tool calling and WikiRacing.

For coding agents, CLM-8B acts as a verifier to select the best solution from multiple candidates. Using fine-tuned heads, the team reports 81.6% accuracy on DeepSWE and 87.6% on Terminal-Bench 2.1 subsets, claiming these are new state-of-the-art verifier results. These figures are based on held-out subsets rather than full leaderboard submissions. The model’s architecture allows for cached state vectors, reducing latency in agent loops from 1.7 ms to 0.6 ms on an RTX 4090.