Stanford and Nvidia researchers released the Contrastive Language Model (CLM) method and the CLM-8B model, which makes decisions by matching state and action representations instead of generating tokens one by one, with reusable cached action representations. In zero-shot tests on computer use, gaming, and tool calling, it ran up to 9x faster than TypeSafe's Jev, scored 95.2% on BFCL v4 (Jev: 99.2%), and completed 26/30 WikiRacing tasks (Jev: 30).