Stanford's CLM: an 8B System 1 model with 9x faster inference, new agentic coding SOTA
StanfordAILab · x · 2026-09-25
A Stanford team (Hazy Research and collaborators) introduced the Contrastive Language Model (CLM), an ultra-fast System 1 model connecting states and actions via contrastive learning:
- CLM-8B, pre-trained on internet-scale data, delivers up to 9x faster inference than Jev with comparable performance on computer use, gaming, and tool calling
- With lightweight fine-tuning it sets new SOTA on agentic coding benchmarks: DeepSWE 81.6% and Terminal-Bench 2.1 87.6%; Jev fails to serve as an effective verifier for these long-horizon tasks
- The team also built efficient training/serving infrastructure by disaggregating states and actions, and established scaling laws across compute, model size, and dataset size
Code, data, and models are released openly.
Related event: Stanford's CLM-8B: 9x Faster Inference, Tops Agentic Coding Benchmarks(4 posts)→
More from coding & agent
- Too many 'X times faster' claims hide extremely narrow benchmarks, says Ruff author — charliermarsh · 2026-09-25
- Ruff author Charlie Marsh: microbenchmark culture is getting worse — charliermarsh · 2026-09-25
- Claude API adds inline tool definitions for adding full tools mid-conversation — EricBuess · 2026-09-25
- Hands-on repo teaches GitHub stacked PRs with gh stack CLI and an AI coding agent workshop — DanWahlin · 2026-09-25
- Energy's AI agent can log in and pay for you, with credentials kept local and hidden from LLMs — emilahlback · 2026-09-25
- OpenRouter launches unified web search server tool with per-provider pricing — jeff_weinstein · 2026-09-25