Contrastive-LM Releases Open CLM-8B Model for Faster Agent Action Scoring

Contrastive-LM has released CLM-8B, marking the debut of the first open model in a new category termed Contrastive Language Models (CLMs). Unlike traditional language models that generate text, CLM-8B is designed to score a set of candidate actions against the current state of an agent, returning probabilities for each action. This new model positions itself as an alternative to proprietary systems like Jev, the System One model developed by TypeSafe AI. CLM-8B is presented as a deployable solution, with its Apache-2.0 licensed head weighing 75 MB. It is capable of running on a single NVIDIA GPU under Linux, utilizing vLLM to serve a Qwen3-8B encoder. The release aims to provide an open-source option for agent action scoring, a critical component in the development of AI agents.

System One models, such as Jev which entered limited early access on September 15, 2026, are characterized by their ability to return typed values with associated probabilities, rather than generating free-form text. CLM-8B targets this same interface, offering a compatible API. The CLM GitHub repository provides access to CLM-8B through a TypeSafe-compatible API, exposing three primary question types: Noul, which returns the probability that a given statement is true; Choice, which selects one option from a declared set with associated probabilities; and Score, which returns an expected level on an ordered rubric. This compatibility allows requests written for TypeSafe's API to be replayed through CLM’s Python client, facilitating adoption and comparison.

The underlying architecture of CLM-8B involves training a state encoder and an action encoder using a bidirectional InfoNCE loss. Each encoder is built upon a frozen Qwen3-8B backbone, augmented with a 20 million parameter trainable projection head. The training process is designed to draw the embedding of the current state closer to the embedding of the action that was actually taken, while simultaneously pushing it away from the embeddings of other candidate actions. During inference, CLM-8B scores each candidate action by calculating the dot product of the state and action embeddings. A subsequent softmax function is applied to these scores to generate the final answer distribution. This design principle allows for the disaggregation of states and actions, which is particularly beneficial in agent loops where the state evolves at each step, while the set of possible actions often remains relatively constant.

The efficiency of CLM-8B is highlighted by its performance metrics. The clm-serve component reserves a dedicated portion of GPU memory, analogous to vLLM’s KV cache, enabling the reuse of cached vectors. This optimization significantly reduces inference latency. For instance, on a single RTX 4090 GPU with three candidate actions, the time to process revisited states drops from 1.7 milliseconds to 0.6 milliseconds. The model card reports that CLM-8B can operate up to 13 times faster than Jev when evaluating approximately 1,000 candidate actions. The training regimen for CLM-8B involves a three-stage process, beginning with pre-training on roughly 60 million Nemotron DQA question-answer pairs, followed by mid-training on an unspecified dataset.