Contrastive Language Models is a System One model trained with a contrastive learning objective that connects states and actions for fast decision-making.
CLM addresses the need for rapid inference in agentic systems by disaggregating state and action representations so their embeddings can be cached and reused independently. The approach trains on question-answer pairs, synthetic hard negatives, and agentic trajectories to learn a scoring function that ranks candidate actions against observed states. Rather than generating text autoregressively, the model answers typed questions about states and ranks free-form candidates directly, enabling both structured and unstructured decision-making through a single primitive: scoring a candidate against a state.
CLM-8B performs comparably to Jev on computer-use, gaming, and tool-calling tasks while achieving up to nine times lower latency, with the largest speedups occurring when candidate action sets are large or actions are reused across states. The tool suits projects requiring fast inference for agent decision-making, particularly those where caching action embeddings across multiple states provides efficiency gains. It ships with a TypeSafe-compatible API, a web playground for interactive exploration, and fine-tuning tutorials for custom data. The model achieves state-of-the-art results as a verifier on agentic coding benchmarks after lightweight fine-tuning.
The project maintains an active community Discord and publishes technical documentation through a blog. Development includes regular updates to the model and serving infrastructure, with examples and benchmarks shipped directly in the repository to support reproducibility and adoption.
