CLM-v0.1-8B

CLM-v0.1-8B, the first “contrastive language model” from the Contrastive-LM research group, was announced on 23 September 2026 and reported by MarkTechPost the same day; its Hugging Face repository was created on 21 September. It does not generate text. It pairs a frozen Qwen3-8B encoder with two small trained projection heads, one for a situation and one for a candidate action, and scores how well each action fits — which lets an AI agent rank its possible next steps, or check a proposed step, in a single pass. MarkTechPost describes the trained part as about 20 million parameters, around 75 MB on top of the base model. The weights are released under Apache 2.0, which allows commercial use. The authors — Jacky Kwok, Hangoo Kang, Tarun Suresh, Jon Saad-Falcon, Marco Pavone, Christopher Ré and Azalia Mirhoseini — report up to nine times lower latency than TypeSafe’s Jev when used without training, with tool-calling success of 95.2 per cent against Jev’s 99.2, and, as a fine-tuned verifier, 81.6 per cent on DeepSWE and 87.6 per cent on Terminal-Bench 2.1 on held-out subsets. Those are the authors’ own results and are not stored as scores. The model card does not state the authors’ affiliations. It is one of several open alternatives to Jev released within a fortnight, which is useful to a buyer mainly as evidence that the approach is not tied to one supplier.