CLM v0.1 8B

CLM v0.1 8B (Contrastive Language Model) is a non-generative model that scores candidate actions or answers against a given state. It consists of two small projection heads, a state head and an action head, on top of a frozen Qwen3-8B encoder, trained with a bidirectional InfoNCE contrastive loss. According to the model card, training used about 60M Nemotron Q&A pairs, about 30M synthetic hard negatives, and about 1M agentic trajectories.

The model ranks free-form candidates such as best-of-N solutions, tool names or next moves, and can answer typed questions (yes/no, choice, score) about a state with probability distributions. States and actions are encoded separately, so action embeddings can be cached; the card claims up to 9× lower latency than "Jev" zero-shot at comparable performance on computer-use, gaming and tool-calling tasks, and 13× with around 1k candidates. Fine-tuned verifier heads are reported at 81.6% on DeepSWE and 87.6% on Terminal-Bench 2.1; these are vendor-reported figures and come from fine-tuned heads, not this checkpoint zero-shot.

It is used through the contrastive-lm Python package, which needs a vLLM-served Qwen3-8B for embeddings and a local clm-serve API with a web playground. Only the heads are trained during fine-tuning. Limitations: it is locked to Qwen3-8B last-token-pooled embeddings, cannot generate text, its probabilities are relative to the supplied candidate set, and it is English-only. A multimodal CLM-35B is announced by the authors.