Contrastive Language Models

Contrastive Language Models is an open-source research framework for fast System One decision models that score actions against states instead of generating long text responses. CLM-8B uses separate state and action encoders trained with contrastive objectives. Candidate embeddings can be cached and reused, reducing latency when actions repeat across requests. The package supports typed yes-or-no, choice, score, and ranking tasks through a TypeSafe-compatible API. It includes inference servers, a browser playground, evaluation tools, preprocessing utilities, and fine-tuning scripts. The reference model uses a Qwen3-8B backbone with lightweight trainable projection heads. It targets routing, verification, tool selection, gaming, and other fast decision workloads.