Stanford and Nvidia’s CLM-8B model runs up to 9x faster than Jev

Stanford and Nvidia just dropped an AI model that makes real-time agent decisions roughly nine times faster than the current leading system. CLM-8B, released on September 23, is the first publicly available Contrastive Language Model, a category that didn’t exist until this paper landed.

The model matches the performance of TypeSafe AI’s proprietary Jev model across multiple benchmarks while cutting latency to a fraction. On the T-Rex game benchmark, CLM-8B clocked 16.5 milliseconds per decision versus Jev’s 149.8 ms.

How contrastive learning changes the game

Instead of generating responses from scratch, it uses contrastive learning to build a shared embedding space where states and actions live side by side. When the model needs to decide what to do next, it scores candidate actions by how closely they match the current state in that embedding space.

The architecture sits on top of a frozen Qwen3-8B backbone. The key innovation is in the projection heads, small trainable modules of roughly 20 million parameters each that learn to map inputs into the contrastive space. The base model’s weights stay locked, which keeps compute costs manageable while the lightweight heads do the heavy lifting for action selection.

Led by researcher Jacky Kwok, the team is calling these “System One” models, borrowing the term from Daniel Kahneman’s framework for fast, intuitive thinking versus slow, deliberate reasoning.

Training at scale, benchmarks that matter

The model went through three distinct training phases. Pre-training used 60 million question-answer pairs to establish baseline understanding. Mid-training introduced 30 million synthetic hard negatives. Post-training refined everything 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 hit 87.6%. Both represent state-of-the-art performance for models in this parameter range.

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

Open weights, single GPU, Apache 2.0

CLM-8B ships with open weights under an Apache 2.0 license. The code and model files are available on GitHub. You can self-host the entire thing on a single Nvidia GPU using vLLM.

A multimodal successor called CLM-35B is already in development, with a targeted release in early October 2026.