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

MarkTechPost · 3 weeks ago · 31

Contrastive-LM released CLM-8B, an open Contrastive Language Model that scores candidate agent actions and returns probabilities instead of generating text. The Apache-2.0 model head is 75 MB and it runs on 1 NVIDIA GPU under Linux using vLLM with a Qwen3-8B encoder. Its cached state/action embeddings and dot-product scoring cut agent-loop latency and, in reported tests, give up to 9× lower latency than Jev while fine-tuned heads improve verifier pass rates on DeepSWE and Terminal-Bench 2.1 subsets.