Stanford and Nvidia have released CLM-8B, an AI model that makes real-time agent decisions roughly nine times faster than the current leading system. The model, made available on September 23, is the first publicly released Contrastive Language Model, a category that did not exist prior to the accompanying paper.

CLM-8B matches the performance of TypeSafe AI's proprietary Jev model across multiple benchmarks while reducing latency to a fraction of previous levels. On the T-Rex game benchmark, CLM-8B recorded 16.5 milliseconds per decision, compared with 149.8 milliseconds for Jev. For agents that chain many decisions into a single run, that per-decision overhead accumulates across every step, which is why the latency figure carries as much weight as the accuracy scores.

How Contrastive Learning Changes the Approach

Rather than generating responses from scratch, CLM-8B applies contrastive learning to construct a shared embedding space in which states and actions coexist. When the model determines its next move, it scores candidate actions according to how closely they match the current state within that space.

The architecture is built on top of a frozen Qwen3-8B backbone. The central innovation lies in the projection heads—small trainable modules of roughly 20 million parameters each that learn to map inputs into the contrastive space. Because the base model's weights remain locked, compute costs stay manageable while the lightweight heads perform the heavy lifting for action selection. Each head amounts to roughly a quarter of one percent of the backbone's eight billion parameters, an asymmetry that shows how little of the network is actually retrained.

Led by researcher Jacky Kwok, the team describes these as "System One" models, a term borrowed from Daniel Kahneman's framework distinguishing fast, intuitive thinking from slow, deliberate reasoning — a framing that places rapid, automatic decision-making rather than extended deliberation at the center of the design.

Large-Scale Training and Benchmark Results

The model was trained across three distinct phases. Pre-training used 60 million question-answer pairs to establish baseline understanding. Mid-training introduced 30 million synthetic hard negatives. Post-training refined the system on 1 million agentic trajectories.

On DeepSWE, a coding benchmark that tests software engineering capabilities, CLM-8B scored 81.6%. On Terminal-Bench 2.1, it reached 87.6%. Both results represent state-of-the-art performance for models in this parameter range, achieved here by openly downloadable weights rather than a closed system.

Beyond coding, the model demonstrated zero-shot performance comparable to Jev on tasks spanning computer use, gaming environments such as Super Mario, and WikiRacing. The speed advantage held across every tested domain, not only the T-Rex benchmark from which the 9x figure derives.

Open Weights Under Apache 2.0

CLM-8B ships with open weights under an Apache 2.0 license, a permissive license that allows commercial use and redistribution with attribution. The code and model files are available on GitHub, and the full system can be self-hosted on a single Nvidia GPU using vLLM — a deployment footprint that keeps hardware requirements within reach of smaller teams, in contrast to the proprietary Jev model it matches.

A multimodal successor, CLM-35B, is already in development, with a targeted release in early October 2026. That release will be the next checkpoint for whether the contrastive approach carries into multimodal settings.