Contrastive-LM Releases CLM-8B: An Open System One Model That Scores Agent Actions Up to 9× Faster Than Jev

Contrastive-LM released CLM-8B, an open model that scores agent actions up to 9× faster than Jev.

Contrastive-LM released CLM-8B, an Apache-2.0 model that scores candidate actions rather than generating text, using a frozen Qwen3-8B encoder and a 75 MB head served with vLLM. Against TypeSafe AI’s proprietary Jev model, zero-shot latency was as much as about 9 times lower, including 16.5 ms versus 149.8 ms on a T-Rex task. Fine-tuned verifier heads reached 81.6% on a 38-task DeepSWE subset and 87.6% on 30 Terminal-Bench 2.1 tasks. Training used roughly 60 million Nemotron DQA pairs, 30 million Gemini 2.5 Flash-Lite hard negatives, and 1 million agent trajectories.

• CLM-8B scores candidate actions with probabilities instead of generating text.
• It uses frozen Qwen3-8B encoders and a 75 MB Apache-2.0 head.
• Zero-shot T-Rex latency is 16.5 ms versus 149.8 ms for Jev.
• Fine-tuned heads reach 81.6% on DeepSWE and 87.6% on Terminal-Bench subsets.
• Training used about 60M DQA pairs, 30M hard negatives, and 1M trajectories.