Contrastive-LM releases open model for scoring AI agent actions

Contrastive-LM has released CLM-8B, an open model that scores possible actions for an AI agent instead of generating text. It compares a description of the current situation with candidate actions, then returns probabilities that software can use to choose a tool or rank proposed solutions. Its Apache-2.0-licensed projection head can be served with a Qwen3-8B encoder on one Nvidia GPU. The team reports lower latency than proprietary rival Jev across its tests, with the cited 9× advantage measured in a game where actions repeat. Its coding-verifier results used additional fine-tuning and small held-out task subsets rather than full leaderboard submissions.