Contrastive-LM Releases Open Agent Action-Scoring Model

Contrastive-LM released an Apache-2.0 Qwen3-8B model that scores candidate agent actions up to about 9× faster than Jev.

Contrastive-LM released an Apache 2.0 System One model, listed on Hugging Face as CLM-v0.1-8B and called CLM-8B by MarkTechPost, that scores candidate actions rather than generating text. It uses a frozen Qwen3-8B encoder with projection heads—the earlier report says two heads, while MarkTechPost describes a 75 MB head served with vLLM—trained with bidirectional InfoNCE on about 60 million Nemotron pairs, 30 million hard negatives, and 1 million agent trajectories, about 91 million examples in total. The sources disagree on details: Hugging Face calls the Nemotron data Q&A pairs and the negatives synthetic, while MarkTechPost says Nemotron DQA pairs and Gemini 2.5 Flash-Lite hard negatives. Zero-shot, it is described as matching TypeSafe AI’s proprietary Jev on computer-use, gaming, and tool use at up to about 9× lower latency; MarkTechPost cites 16.5 ms versus 149.8 ms on a T-Rex task, and the Hugging Face note also says it is 13× faster with about 1,000 candidates. Fine-tuned heads reach 81.6% on DeepSWE, specified as a 38-task subset, and 87.6% on Terminal-Bench 2.1, specified as 30 tasks. According to the Hugging Face report, a multimodal CLM-35B is planned for early October.

• Hugging Face (2026-09-21) lists the release as CLM-v0.1-8B, trending #30; MarkTechPost (2026-09-24) calls it CLM-8B.
• Apache 2.0 System One model scores given candidate actions and does not generate text, using a frozen Qwen3-8B encoder; MarkTechPost says a 75 MB head is served with vLLM.
• Trained with bidirectional InfoNCE on about 60 million Nemotron pairs, 30 million hard negatives, and 1 million agent trajectories (about 91 million examples). Sources differ on Q&A versus DQA pairs and on whether negatives are only…
• Zero-shot latency versus TypeSafe AI’s proprietary Jev is up to about 9× lower; MarkTechPost cites 16.5 ms versus 149.8 ms on a T-Rex task. The Hugging Face report also says it matches Jev on computer-use, gaming, and tool use and is 13×…
• Fine-tuned heads reach 81.6% on DeepSWE (a 38-task subset, per MarkTechPost) and 87.6% on Terminal-Bench 2.1 (30 tasks, per MarkTechPost).
• The Hugging Face report says a multimodal CLM-35B is planned for early October.