Contrastive Language Models: CLM-8B picks actions by embedding similarity, matching Jev at up to 9x lower latency

Contrastive Language Models project·medium signalXBlueskyLinkedInCopy link

The CLM team trains separate state and action encoders with an InfoNCE objective and scores candidate actions by embedding alignment, so it generates no text. CLM-8B was pre-trained on 60M Nemotron Q&A pairs and post-trained on 1M agentic trajectories. The team reports Jev-level zero-shot results on computer use, games and tool calling, and after fine-tuning 81.6% on DeepSWE and 87.6% on Terminal-Bench 2.1. Because action embeddings can be cached, this is the first open-weights System One model that is cheap to self-host, though the coding numbers need independent replication.