Stanford researchers have released CLM-8B, an open model designed to make fast, structured decisions for AI agents instead of generating text one token at a time.
The model is part of a new approach called Contrastive Language Models, or CLMs, which connects an agent’s current state with possible actions and scores how well those actions fit the state.
The project is led by Jacky Kwok, Hangoo Kang, Tarun Suresh, Jon Saad-Falcon, Marco Pavone, Christopher Ré and Azalia Mirhoseini. The team describes CLM as a System One model for fast and generalizable decision-making.
CLM-8B is built on a frozen Qwen3-8B encoder with two trainable projection heads: one for states and one for actions. Each projection head has about 20 million parameters.
The model uses a bidirectional InfoNCE contrastive-learning objective. During training, matching state-action pairs are pulled together in representation space while incorrect alternatives are pushed apart.
At deployment, CLM encodes the current state and candidate actions separately, compares their embeddings and selects the highest-scoring action.
The architecture also allows state and action embeddings to be cached independently. This means action representations can be computed once and reused across multiple states, reducing repeated computation when the same candidate actions appear throughout an agent workflow.
The released system supports three decision types: Noul, which returns the probability that a statement is true; Choice, which selects from predefined options; and Score, which places an item on an ordered scale.
CLM can also rank candidate sets directly, including possible tool calls, next moves, retrieval candidates and multiple generated answers.
The model was trained in three stages. The first used about 60 million Nemotron question-and-answer pairs, with each question treated as a state and its answer treated as the corresponding action.
The second stage used about 30 million synthetic hard negatives generated with Gemini 2.5 Flash-Lite. These examples were designed to provide answers that were semantically similar to the correct answer but still incorrect.
The final stage used about 1 million agentic trajectories from the Agent Data Protocol dataset, together with terminal traces from Endless-Terminals and LiteCoder-Terminal-SFT.
During post-training, 40% of the mixture consisted of Nemotron DQA replay and 60% consisted of agentic trajectories. The project reports that removing this replay reduced performance on the Nemotron hard-negative evaluation.
On a held-out set of about 100,000 questions with one correct answer and 10 hard negatives per question, pre-training alone reached 52.1% top-1 accuracy. Adding a short hard-negative training stage raised that figure to 69.2%.
Training with hard negatives from the beginning instead peaked at 62.4% before overfitting, leading the researchers to use broad pre-training before the harder discrimination stage.
The released CLM-8B system is available through the project’s open-source GitHub repository, while the model weights are hosted on the CLM-8B model page on Hugging Face.
The code and CLM-8B weights are released under the Apache 2.0 license. The underlying Qwen3-8B model is also listed under Apache 2.0.
The published serving setup uses vLLM to run Qwen3-8B as a pooling model and a separate CLM server to handle the decision layer. The reference setup uses a maximum model length of 2,048 tokens and a CLM head file of about 75 MB.
That 75 MB figure refers to the CLM projection-head checkpoint, not the complete 8-billion-parameter model. The Qwen3-8B backbone is still required.
The project requires Python 3.10 or later, Linux and an NVIDIA GPU. Its server also provides a browser-based playground and an API compatible with the structure used by TypeSafe’s System One interface.
In zero-shot evaluations covering computer-use, gaming and tool-calling tasks, the researchers report that CLM-8B performs on par with Jev, TypeSafe AI’s System One model, while reaching up to 9 times lower latency.
The largest gains were reported when candidate actions were numerous or repeatedly reused across different states. The project’s examples include the T-Rex game and WikiRacing.
The model card also reports that when the system handles about 1,000 candidate actions, CLM can be about 13 times faster than Jev, reflecting the benefit of reusable action embeddings.
The researchers also tested CLM as a verifier for coding agents. In that setup, a larger model first generates several candidate solutions and CLM scores the candidates before selecting one.
For DeepSWE, the candidate solutions were generated with Opus 5. For Terminal-Bench 2.1, the candidates were generated with Fable 5.
Using lightweight fine-tuning, CLM reached 81.6% on 38 held-out DeepSWE tasks and 87.6% on 30 held-out Terminal-Bench 2.1 tasks.
The project reports that CLM ran between 4.1 and 5.7 times faster than Jev in those verifier evaluations. It also reports that Jev failed to serve as a verifier for the long-horizon tasks tested, scoring below the pass-at-one baseline used in the evaluation.
The DeepSWE result corresponds to 31 successful tasks out of 38 in the released evaluation materials.
The project’s claims about Terminal-Bench are based on Terminal-Bench 2.1, a revision that fixed issues in 28 of the 89 tasks from Terminal-Bench 2.0.
CLM’s verifier results require fine-tuning. The public CLM-v0.1-8B checkpoint itself does not achieve the reported 81.6% and 87.6% results without additional task-specific training.
The model also does not generate text. It only scores candidates supplied to it, and its output probabilities are relative to the candidate set presented in a particular request.
The released checkpoint is tied to the Qwen3-8B encoder and its last-token pooling configuration. The model card states that the projection heads require that encoder and pooling setup.
The project says its approach can be used for tool routing, best-of-N ranking, retrieval shortlisting, typed decisions, computer-use actions and agent verification.
The architecture is based on treating decision-making as a matching problem between a state and a closed set of possible actions rather than asking a large language model to generate each decision sequentially.
That design also places a limit on the system: CLM can only choose among the candidates provided to it. The quality and coverage of those candidates therefore remain part of the overall agent design.
The researchers report scaling relationships in which the InfoNCE loss decreases as training compute, dataset size, projection-head size and encoder size increase. Their experiments indicate that encoder size produces the strongest gains among the tested scaling dimensions.
They also report that the optimal projection-head size grows almost linearly with the number of training tokens, with a fitted relationship of approximately N* ∝ D1.02 and roughly 310 training tokens per parameter.
CLM was released shortly after TypeSafe AI introduced Jev and its System One model architecture on September 15, 2026.
TypeSafe describes Jev as a model built for structured decisions rather than string generation, with outputs designed to be consumed directly by software. CLM follows the same broader decision-oriented model category while making its state-action representation and caching mechanism available as an open implementation.
The project’s training recipe relies heavily on NVIDIA’s Nemotron data. NVIDIA describes Nemotron as a family of open models and datasets intended for building specialized AI agents, with training data, model weights and technical recipes made available to developers.
CLM’s release does not replace the role of a general-purpose language model. Instead, its design separates generation and reasoning from repeated decision steps that can be reduced to scoring and ranking.
The researchers plan to continue scaling the model with larger backbones, more pre-training data, additional hard-negative mining and more agentic post-training.
The roadmap also includes vision and multimodal support for robotics and computer-use applications.
The model card says a larger CLM-35B multimodal model, trained with more data, compute and parameters, is planned for early October.
The project is inviting developers to integrate CLM into their own agents and benchmark the model against other decision systems.
Discover more from Aree Blog
Subscribe now to keep reading and get access to the full archive.
