CLM-v0.1-8B: 9x Faster LLM Inference, Achieving SOTA in Agent Verification

This is a System One model that places two projection heads on top of the Qwen3-8B encoder to connect states and actions, enabling immediate decision-making without complex reasoning. Instead of generating answers like a generative LLM, it scores and ranks the optimal choices from a given set of candidates, significantly reducing latency.

It demonstrates performance equivalent to the Jev model on agent tasks such as computer use, gaming, and tool calling, while reducing latency by up to 9x. Notably, fine-tuned verifiers achieved SOTA scores of 81.6% and 87.6% on the DeepSWE and Terminal-Bench 2.1 benchmarks, respectively, proving processing speeds 4–6x faster than existing solutions.

By adopting a caching structure that separately encodes states and actions, it is 13x faster than Jev when evaluating approximately 1,000 candidates. It is suitable for use as a lightweight inference engine in agent workflows requiring repetitive decision-making, such as customer inquiry classification or next-action selection.