Contrastive Language Model debuts with claims of up to 9× faster inference than Jev
The CLM-8B announcement claims performance comparable to Jev across computer-use, gaming and tool-calling tasks, plus state-of-the-art coding benchmark results after lightweight fine-tuning.
TLDR
CLM’s developers describe a model trained to connect situations and actions through contrastive learning. They claim CLM-8B delivers up to 9× faster inference than Jev with comparable performance across computer-use, gaming and tool-calling tasks. After lightweight fine-tuning, they report scores of 81.6% on DeepSWE and 87.6% on Terminal-Bench 2.1, calling both state-of-the-art results. Their setup caches representations of situations and actions separately, which they say reduces latency when situations change but the available actions stay fixed.
