Contrastive-LM Releases CLM-8B: Open Action-Scoring Model Up to 9x Faster Than Jev

Contrastive-LM has released CLM-8B, the first open model in a new class called Contrastive Language Models. Instead of generating text, it scores candidate actions and returns probabilities, running up to 9x faster than TypeSafe AI's proprietary Jev on T-Rex. The Apache-2.0 head weighs 75 MB and runs on a single Nvidia GPU.

• CLM-8B scores candidate actions instead of generating text
• Up to 9x faster than Jev on T-Rex, 4.1–5.7x on DeepSWE and Terminal-Bench 2.1
• Top-1 accuracy: 52.1% after pre-training, 69.2% after mid-training
• Apache-2.0 head weighs 75 MB, self-hosted on 1 Nvidia GPU via vLLM