{"version":1,"type":"story","url":"https://digestai.news/story/contrastive-lm-releases-clm-8b-an-open-model-scoring-actions-up-to-9x","json":"https://digestai.news/story/contrastive-lm-releases-clm-8b-an-open-model-scoring-actions-up-to-9x.json","markdown":"https://digestai.news/story/contrastive-lm-releases-clm-8b-an-open-model-scoring-actions-up-to-9x.md","slug":"contrastive-lm-releases-clm-8b-an-open-model-scoring-actions-up-to-9x","headline":"Contrastive-LM releases CLM-8B, an open model scoring actions up to 9x faster than Jev","summary":"Contrastive-LM has released CLM-8B, described as the first open model in the Contrastive Language Models (CLMs) class. Unlike traditional generative models, CLM-8B does not generate text; instead, it scores candidate actions against a current state and returns probabilities. The model uses a frozen Qwen3-8B backbone with a small trainable projection head, allowing it to run on a single NVIDIA GPU under Linux. The Apache-2.0 licensed head weighs only 75 MB, making it lightweight for deployment.\n\nThe release targets the same interface as Jev, a proprietary System One model from TypeSafe AI that entered limited early access on 15 September 2026. CLM-8B exposes three question types: Noul (truth probability), Choice (option selection), and Score (rubric evaluation). In zero-shot tests, the research team reports CLM-8B runs up to 9 times faster than Jev, particularly in tasks with repeated actions like the T-Rex game. While it matches Jev on some tasks, it trails on tool calling and WikiRacing.\n\nFor coding agents, CLM-8B acts as a verifier to select the best solution from multiple candidates. Using fine-tuned heads, the team reports 81.6% accuracy on DeepSWE and 87.6% on Terminal-Bench 2.1 subsets, claiming these are new state-of-the-art verifier results. These figures are based on held-out subsets rather than full leaderboard submissions. The model’s architecture allows for cached state vectors, reducing latency in agent loops from 1.7 ms to 0.6 ms on an RTX 4090.","keyPoints":["CLM-8B is an open-source model that scores actions rather than generating text, using a Qwen3-8B backbone.","The model runs up to 9 times faster than TypeSafe AI's Jev in zero-shot tests on specific tasks.","Fine-tuned CLM heads achieve 81.6% on DeepSWE and 87.6% on Terminal-Bench 2.1 subsets as a verifier."],"whyItMatters":"CLM-8B offers a low-cost, open alternative for agent verification and decision-making, potentially reducing latency and compute costs for developers building autonomous AI systems compared to proprietary options.","category":{"slug":"models","name":"Generative AI & Models","url":"https://digestai.news/category/models"},"entities":{"companies":["Contrastive-LM","TypeSafe AI","NVIDIA"],"models":["CLM-8B","Jev","Qwen3-8B","Gemini 2.5 Flash-Lite","Opus 5","Fable 5"],"people":[]},"firstPublishedAt":"2026-09-24T05:27:41Z","updatedAt":"2026-09-24T05:27:41Z","sourceCount":2,"hasPrimarySource":false,"sources":[{"outlet":"MarkTechPost","title":"Contrastive-LM Releases CLM-8B: An Open System One Model That Scores Agent Actions Up to 9× Faster Than Jev","url":"https://marktechpost.com/2026/09/23/contrastive-lm-releases-clm-8b-an-open-system-one-model-that-scores-agent-actions-up-to-9x-faster-than-jev","publishedAt":"2026-09-24T05:27:41Z","type":"press","primary":false,"lead":true},{"outlet":"contrastive-lm.notion.site","title":"Contrastive Language Models","url":"https://contrastive-lm.notion.site/","publishedAt":"2026-09-24T04:47:11Z","type":"press","primary":false,"lead":false}],"sourceNotes":null,"discussions":[{"site":"Reddit","url":"https://www.reddit.com/r/LocalLLaMA/comments/1woscea/contrastive_language_models/","points":null}],"thread":{"title":"Open-Source LLM Tools Accelerate Real-World AI","url":"https://digestai.news/thread/openarch-launches-pytorch-repo-for-readable-llm-architecture-implementations","storyCount":4},"cite":{"text":"Digest AI, \"Contrastive-LM releases CLM-8B, an open model scoring actions up to 9x faster than Jev\", 24 September 2026, https://digestai.news/story/contrastive-lm-releases-clm-8b-an-open-model-scoring-actions-up-to-9x","publisher":"Digest AI","title":"Contrastive-LM releases CLM-8B, an open model scoring actions up to 9x faster than Jev","datePublished":"2026-09-24T05:27:41Z","url":"https://digestai.news/story/contrastive-lm-releases-clm-8b-an-open-model-scoring-actions-up-to-9x"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}