Stanford and Nvidia released CLM-8B on September 23, the first publicly available contrastive language model. The system makes decisions roughly nine times faster than TypeSafe AI’s proprietary Jev model, cutting latency from 149.8 milliseconds to 16.5 milliseconds on the T-Rex benchmark. The architecture uses a frozen Qwen3-8B backbone with trainable projection heads of roughly 20 million parameters each. The model scores 81.6% on DeepSWE and 87.6% on Terminal-Bench 2.1, setting state-of-the-art performance for its parameter range. Released under an Apache 2.0 license, it can run on a single Nvidia GPU via vLLM. A multimodal successor called CLM-35B is targeted for October 2026.
Source: Read the original article
