Contrastive-LM launches open CLM-8B for scoring agent actions
Contrastive-LM has released CLM-8B, described as the first open model in the Contrastive Language Model class. Rather than generating text, it evaluates candidate actions against a current state and produces probabilities. The system is designed to match the interface of Jev, TypeSafe AI’s proprietary System One model, and its repository provides a TypeSafe-compatible API with Noul, Choice and Score request types.
CLM-8B uses frozen Qwen3-8B backbones with two trainable 20M-parameter projection heads. A bidirectional InfoNCE objective brings the representation of a state closer to the action taken and separates it from alternatives. At inference, dot-product scores are converted into a distribution with softmax. Cached state and action vectors can reduce repeated-state latency from 1.7 ms to 0.6 ms on an RTX 4090 with three actions.
The model card reports a 13× speed advantage over Jev with about 1,000 candidates. In zero-shot tests, it matched Jev on T-Rex and Super Mario, but lagged on tool calling and WikiRacing. Fine-tuned heads reached 81.6% on a DeepSWE subset and 87.6% on a Terminal-Bench 2.1 subset. The Apache-2.0 head is 75 MB and can run on one NVIDIA GPU under Linux.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.