🚀 Stanford and NVIDIA have released CLM-8B, an open-source AI model that makes decisions up to 9x faster than Jev without generating a single word. While most LLMs are designed to produce text, Contrastive Language Models (CLMs) are built for something different: helping AI agents choose the right action as quickly as possible. Instead of generating an explanation, CLM receives the current situation and a set of possible actions, then assigns probabilities to each option. It can handle: ▪️ Binary decisions: True/false, yes/no, approve/reject. ▪️ Multiple-choice decisions: Select the best action from a list of alternatives. ▪️ Ordinal scoring: Rank options on an ordered scale. ▪️ Agent verification: Evaluate several solutions generated by a coding agent and select the most promising one. How does CLM-8B work? The architecture is based on Qwen3-8B, with two lightweight projection heads trained on top of a frozen backbone. One encodes the current state, while the other encodes possible actions. The model compares their vector representations and converts similarity scores into probabilities. Here's the clever part: action caching. If an agent repeatedly chooses between the same actions, their embeddings can be calculated once and reused. That eliminates redundant computation and makes repeated decisions significantly faster. Benchmark results reported by the researchers: ▪️ Up to 9x lower latency than Jev on computer-use, gaming, and tool-calling tasks, with broadly competitive performance. ▪️ Up to 13x faster when choosing between approximately 1,000 cached candidate actions. ▪️ DeepSWE: Fine-tuned CLM improved coding task success from 73.7% to 81.6% by selecting among multiple generated solutions. ▪️ Terminal-Bench 2.1: Success increased from 84.0% to 87.6% using the same verification approach. For comparison, Jev achieved 71.1% on DeepSWE and 83.1% on Terminal-Bench 2.1 in the researchers' verifier evaluations. The verifier comparisons used small held-out sets of 38 and 30 tasks, and CLM received task-specific fine-tuning, so these results shouldn't be interpreted as definitive proof of general superiority. The model is available under Apache 2.0, with public weights, inference code, and fine-tuning tools. Running the reference implementation requires Linux, an NVIDIA GPU, and Qwen3-8B served through vLLM. The bigger idea is fascinating: not every AI agent decision needs a massive reasoning model generating hundreds of tokens. Sometimes an agent already knows its options. It just needs to pick the best one. And that's exactly where specialized decision models could make autonomous AI systems faster, cheaper, and more scalable. Research and documentation: contrastive-lm.notion.site/ Models on Hugging Face: huggingface.co/Contrastive-L… GitHub: github.com/Contrastive-LM/CL… #AIAgents #OpenSourceAI #MachineLearning