Contrastive Language Models
CLM-8B is a System One model trained with contrastive learning to score candidate actions against a state, matching a larger reference model's accuracy on computer-use, gaming and tool-calling tasks at up to 9x lower latency.
Contrastive Language Models (CLMs) are a class of System One model trained with a contrastive learning objective connecting states and actions. CLM-8B is pre-trained on 60 million Q&A pairs, mid-trained on 30 million synthetic hard negatives, and post-trained on 1 million agentic trajectories, and performs on par with a larger reference model across computer-use, gaming and tool-calling tasks at up to 9 times lower latency. With lightweight fine-tuning it sets a new state of the art as a verifier on agentic coding benchmarks.
States and actions are disaggregated so their embeddings can be cached and reused independently, keeping training and serving cheap. The model is served behind a typed API (clm-serve) alongside a vLLM-served embedding model, and exposes typed questions about a state, or ranking of free-form candidates such as best-of-N answers or tool names.
A built-in web playground lets a developer write a state, add typed questions, see the answer distribution, and copy the equivalent curl or Python request; a Rank tab ranks arbitrary candidate sets.