Contrastive-LM releases CLM-8B, an open model that scores actions up to 9× faster than Jev
Contrastive-LM has released CLM-8B, the first open Contrastive Language Model, which scores candidate actions instead of generating text. The model uses a frozen Qwen3-8B backbone with trainable projection heads and is designed to work with TypeSafe AI's proprietary Jev model interface. It runs on a single NVIDIA GPU under Linux and is served via vLLM.
Key points
- CLM-8B is an open Contrastive Language Model that scores actions instead of generating text
- It runs up to 9× faster than Jev in zero-shot tests, with latency down to 0.6 ms on RTX 4090
- Fine-tuned heads reach 81.6% on DeepSWE and 87.6% on Terminal-Bench 2.1 subsets
In zero-shot tests, CLM-8B runs up to 9× faster than Jev, with latency dropping from 1.7 ms to 0.6 ms on an RTX 4090 for revisited states. The model card reports 13× speedup with ~1,000 candidates. Fine-tuned versions achieve 81.6% on DeepSWE and 87.6% on Terminal-Bench 2.1 subsets. The Apache-2.0 licensed head is 75 MB and self-hostable.
Training involved three stages: pre-training on 60M Nemotron DQA pairs, mid-training on 30M synthetic hard negatives from Gemini 2.5 Flash-Lite, and post-training on 1M agent trajectories. On held-out questions, pre-training reached 52.1% top-1 accuracy, rising to 69.2% after mid-training. The model is positioned as a verifier for coding agents, outperforming Jev in latency while matching or trailing in accuracy depending on the task.
Model page: Jev →
The story so far
4 episodes →- Contrastive-LM releases CLM-8B, an open model that scores actions up to 9× faster than Jevthis story
Contrastive-LM Releases CLM-8B: An Open System One Model That Scores Agent Actions Up to 9× Faster Than Jev
MarkTechPost · 24 September 2026
Loading the full article…
This text was published by MarkTechPost and written by Michal Sutter. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
2sources- Reddit discussionreddit.com
- Contrastive Language ModelsPress · contrastive-lm.notion.site ·
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Apiiro finds 1,122 public GitHub files with secrets saved by Claude Code · 1 src
- Thinking Machines releases Inkling 2 Small, a 594B MoE model with 17B active parameters · 8 src
- OpenAI releases GPT-6 Cyber, a Critical-rated model that outperforms predecessors in vulnerability detection and alignment · 25 src
- Google releases Gemini Omni 1.2 Flash with 15-second video generation and 30% price cut to $0.07 per second · 10 src
- US accuses Chinese executive of allegedly sharing Nvidia chip info with Beijing · 11 src
Comments
via GitHub Discussions