Contrastive-LM Debuts CLM-8B to Score Agent Actions 13× Faster Than Proprietary Models

Contrastive-LM launched CLM-8B, introducing Contrastive Language Models that return action probabilities instead of generating text. The architecture uses a frozen Qwen3-8B backbone with a 75 MB Apache-2.0 projection head that runs on a single NVIDIA GPU. By reserving GPU memory to reuse cached vectors, revisited states drop from 1.7 ms to 0.6 ms on an RTX 4090 with three actions. CLM-8B runs up to 13× faster than TypeSafe AI's Jev baseline at 1,000 candidates and 9× faster on the T-Rex game. On held-out DeepSWE and Terminal-Bench 2.1 tasks, CLM operates 4.1× to 5.7× faster than Jev, achieving new state-of-the-art verifier results.