Stanford and NVIDIA Open-Source CLM-8B, an Open Rival to Jev
Stanford and NVIDIA have released CLM-8B, the first open model in the Contrastive Language Models class and an open counterpart to Jev, which TypeSafe AI introduced on 15 September. Given a current state and a set of candidate actions, it returns a calibrated probability for each, and it never generates text: it can judge whether a claim is true, pick among options, or score on an ordered scale.
The architecture is Qwen3-8B with two small projection heads trained on top, one for state and one for actions. State and actions are encoded separately, the model compares their vectors, and the closeness becomes a probability. Vectors for unchanged actions can be precomputed and reused, which is where the speed on multi-step tasks comes from.
The release reports that, without fine-tuning, CLM-8B matches Jev on computer control, games and tool calling while answering up to 9x faster, and up to 13x with vector caching when choosing from about a thousand options. Fine-tuned as a verifier that picks the best of several coding-agent attempts, it lifts DeepSWE from 73.7% to 81.6% and Terminal-Bench 2.1 from 84% to 87.6%, against 71.1% and 83.1% for Jev in the same role.
CLM-8B is Apache 2.0 licensed and runs on Linux with an NVIDIA GPU and Qwen3-8B served through vLLM.
Related: TypeSafe AI Introduces System One Models and Jev, Microsoft Releases Decision-1, a Model That Only Makes Decisions