CLM-v0.1-8B Released, 13x Faster Decision-Making Compared to Jev

On September 23, 2026, CLM (Contrastive Language Model) v0.1-8B was released, which vectorizes states and actions to quickly select the optimal action based on similarity. This model adopts a System One decision-making structure that returns the highest score among candidates instead of free generation.

Architecture and Training

CLM-v0.1-8B uses Qwen3-8B as an encoder with fixed weights, and only a small Projection Head for state and action embeddings was additionally trained. Through bidirectional InfoNCE Loss, correct pairs are optimized to be close, while other combinations within the batch are pushed apart. Training data includes Nemotron DQA, synthetic hard negatives, and agent trajectories, with hard negative training improving top-1 accuracy by approximately 7%p.

Performance and Speed

By pre-computing and caching action vectors, inference speed is achieved approximately 13x faster than Jev when comparing 1,000 candidates. In zero-shot evaluation, the latency for T-Rex is 16.5ms, approximately 9x lower than Jev (149.8ms), while maintaining a similar success rate. In coding verifier evaluation, the success rate for DeepSWE was 81.6%, surpassing Jev (71.1%).

Deployment and Future Plans

Released under the Apache 2.0 license, it is deployed via vLLM's Qwen3-8B pooling mode and the CLM API server. The head download size is 75MB. Training for a multimodal CLM-35B is underway, with a planned release in early October 2026.