Researchers from Stanford University and NVIDIA have introduced Contrastive Language Models (CLMs), a new class of System One model designed for fast and generalizable decision-making. CLM-8B, the released model, was pre-trained on 60M Nemotron Q&A pairs, mid-trained on 30M synthetic hard negatives, and post-trained on 1M agentic trajectories. The model delivers performance comparable to Jev across computer-use, gaming, and tool-calling tasks while achieving up to 9x lower latency.

A key innovation involves disaggregating states and actions, allowing embeddings to be cached and reused independently for ultra-efficient training and serving infrastructure. Researchers established scaling laws demonstrating that test contrastive loss decreases predictably as a power law in training compute, model size, and dataset size. With lightweight fine-tuning, CLM-8B sets a new state-of-the-art on challenging agentic coding benchmarks, including Deep SWE at 81.6% and Terminal Bench 2.1 at 87.6%.

Performance evaluations across computer-use, gaming, and tool-calling tasks show CLM performs on par with existing models while running significantly faster, with speedups most pronounced when candidate actions are large or frequently reused across states. The team released the model on GitHub for public access.