
CLM-8B launches open, fast agent decision model
Stanford and NVIDIA researchers released CLM-8B, an Apache 2.0 contrastive model that scores candidate actions through shared state-action embeddings instead of generating text. Its independently cached action embeddings deliver up to 9× lower latency than Jev on selected agent, gaming, and tool-calling tasks.
CLM-8B makes a compelling case for separating fast decisions from expensive reasoning, though its headline speed advantage depends heavily on reusable action sets.
- –Frozen Qwen3-8B embeddings and lightweight trainable heads make the system practical to fine-tune and self-host
- –Best suited to tool routing, candidate ranking, verification, and other bounded-choice workflows
- –Reported speedups range from 1.6× to 9×, with the largest gains coming from cached or numerous candidate actions
- –Accuracy trails Jev on BFCL tool calling and WikiRacing, so latency gains require task-specific validation
- –Fine-tuned verifier results on DeepSWE and Terminal-Bench suggest a promising selection layer for multi-model coding agents
DISCOVERED
1h ago
2026-09-27
PUBLISHED
1h ago
2026-09-27
RELEVANCE
AUTHOR
Prompt Engineering