NVIDIA and Stanford Release CLM-8B: 9x Faster Inference, New Agentic Coding SOTA
NVIDIA and Stanford (Hazy Research and collaborators) released the Contrastive Language Model (CLM), a System 1 model that links states to actions through a contrastive learning objective. The flagship CLM-8B, pretrained on internet-scale data, matches Jev zero-shot while offering up to 9x faster inference, and after light fine-tuning it surpasses DeepSWE's 81.6% and tops Terminal-Bench, setting a new SOTA on agentic coding benchmarks.
Confirmed
- CLM comes from an NVIDIA–Stanford collaboration and uses a System 1 architecture: states and actions are connected via contrastive learning rather than token-by-token generation
- According to @wschroll's detailed write-up, CLM generates no text; it handles high-frequency small decisions inside AI systems—tool selection, patch ranking, request routing, next-action choice—treating decisions as retrieval problems
- CLM-8B is pretrained on internet-scale data, with zero-shot performance on par with Jev and inference up to 9x faster
- With light fine-tuning, CLM-8B breaks DeepSWE's 81.6% and tops Terminal-Bench
Why it matters
- CLM demonstrates a route distinct from "ever-larger generative models": peeling high-frequency small decisions out of the generative pipeline and replacing them with retrieval-style contrastive learning yields large speedups at comparable performance
- Following the release of the small model Jev, the community has built applications such as live trading bots deciding whether to place orders every 300ms; CLM's 9x inference speedup could further cut agent latency and cost for tool calling, patch ranking, and request routing
2026-09-24 ~ 2026-09-26 · 6 related posts
Primary sources
- CLM-8B: contrastive System One model claims 9x faster inference, agentic SOTA — ChengleiSi · 2026-09-24
- CLM-8B hits SOTA 81.6% on DeepSWE with light finetuning, up to 9x faster inference — anshulkundaje · 2026-09-24
- Contrastive Language Model is 9x faster than Jev, sets SOTA on agentic coding benchmarks — iamrobotbear · 2026-09-24
- [source] Stanford's CLM: an 8B System 1 model with 9x faster inference, new agentic coding SOTA — StanfordAILab · 2026-09-25
- NVIDIA and Stanford's CLM Treats Decision-Making as Retrieval, Up to 9x Faster — wschroll · 2026-09-26
1 near-duplicate retellings: wschroll