By No Code MBA · Facts checked
- Provider
- Stanford and NVIDIA
- Announcement date
- Status at last check
- Generally available
- Model family
- CLM
- Capabilities
- Text
What changed
CLM-8B is a contrastive language model that scores candidate actions against a state instead of generating tokens. It pairs a frozen Qwen3-8B backbone with trainable projection heads of about 20M parameters each, trained on roughly 60M question-answer pairs, 30M synthetic hard negatives and 1M agent trajectories. The authors report results on par with Jev on computer use, gaming and tool-calling tasks with up to 9x lower latency, and 87.6% on Terminal-Bench 2.1 as a fine-tuned verifier.
Access and availability
Install with pip install contrastive-lm, serve the Qwen3-8B encoder with vLLM, then run clm-serve for HTTP endpoints such as /v1/rank. Code and weights (Contrastive-LM/CLM-v0.1-8B on Hugging Face) are Apache 2.0.
Official access documentationWhat you can build
Use CLM-8B as a fast action selector or verifier in agent pipelines where the model chooses between a fixed set of tools or candidate outputs, and cache action embeddings to cut latency on repeated decisions.
Limitations
Self-hosted only, with states truncated at 2048 tokens by default. The strongest verifier results rely on fine-tuning, candidates from other models and small held-out sets, and vision support and larger backbones are roadmap items. Claims are the authors' own.