# Contrastive-LM releases CLM-8B, an open model scoring actions up to 9x faster than Jev

Digest AI · Generative AI & Models · published 2026-09-24T05:27:41Z

Canonical: https://digestai.news/story/contrastive-lm-releases-clm-8b-an-open-model-scoring-actions-up-to-9x

## 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.

The 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.

For 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.

## Key points

- 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.

## Why it matters

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.

## Sources

1. [Contrastive-LM Releases CLM-8B: An Open System One Model That Scores Agent Actions Up to 9× Faster Than Jev](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) (MarkTechPost, 2026-09-24)
2. [Contrastive Language Models](https://contrastive-lm.notion.site/) (contrastive-lm.notion.site, 2026-09-24)

Part of the developing story: [Open-Source LLM Tools Accelerate Real-World AI](https://digestai.news/thread/openarch-launches-pytorch-repo-for-readable-llm-architecture-implementations) (4 stories)

## Cite

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

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. 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
JSON: https://digestai.news/story/contrastive-lm-releases-clm-8b-an-open-model-scoring-actions-up-to-9x.json