CLM-8B: No Text, Just Decisions

🔥 What happened Contrastive-LM released CLM-8B, the first open model in a new class called Contrastive Language Models. It doesn't generate text—it scores candidate actions and returns probabilities. The Apache-2.0 head weighs just 75 MB and runs on a single NVIDIA GPU. 💡 Why it matters CLM-8B is up to 9× faster than Jev, TypeSafe AI's proprietary System One model, in zero-shot tests. On tool calling, it hits 95.2% success vs. Jev's 99.2%, but at 76.8 ms instead of 125.5 ms. Fine-tuned, it reaches 81.6% on DeepSWE and 87.6% on Terminal-Bench 2.1—new SOTA verifier results. ⚡ Our take A model that doesn't talk but only judges is exactly what agent loops need. Anyone still betting on text generation as the universal solution is leaving latency and precision on the table.