🔬Stanford, Nvidia's CLM-8B Runs Agent Decisions 9x Faster
TL;DR
Stanford and Nvidia open-sourced CLM-8B, an 8-billion-parameter model that caches reusable agent actions and beats TypeSafe's Jev by up to 9x on latency. It trades a little tool-calling accuracy for major speed gains, built on a frozen Qwen3-8B backbone and 60 million QA pairs.
Stanford and Nvidia open-sourced CLM-8B, an 8-billion-parameter model that caches reusable agent actions and beats TypeSafe's Jev by up to 9x on latency. It trades a little tool-calling accuracy for major speed gains, built on a frozen Qwen3-8B backbone and 60 million QA pairs.
Key Points
CLM-8B is an 8B-parameter open model built on a frozen Qwen3-8B backbone, trained on about 60 million question-answer pairs and 1 million agent trajectories
Runs up to 9x faster than TypeSafe's Jev in zero-shot tests, with 4.1-5.7x lower verifier latency
Scored 87.6% on Terminal-Bench 2.1 versus Jev's 83.1%, and 81.6% on DeepSWE verification versus Jev's 71.1%
Trails Jev on BFCL v4 tool-calling (95.2% vs. 99.2%) and WikiRacing (26/30 vs. 30/30)
Project lead Jacky Kwok: "Use large reasoning models to generate and reason, and use CLMs to cheaply select, verify, and monitor their outputs"
Why It Matters
Pairing a big reasoning model with a cheap, fast verifier like CLM-8B is a cost pattern every team running agents at scale should be watching.
Quick Facts
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.