SenseTime open-sources a 9B model that matches 27B performance

How small models can match big ones is a hot topic in the AI industry. SenseTime, a Chinese AI company, has open-sourced SenseNova Loop-9B-Preview, a 9B model that performs like a 27B one.

The model uses a Looped Transformer design. It keeps 4 decoder layers at each end and turns the middle layers into shared looped computing modules, deepening its effective computing power through repeated iterations. It also adds mHC to boost feature interaction.

In tests, the model beat similar models in coding, office tasks, and agent work. It also topped Qwen3.5-27B in benchmarks like WorkBuddy Bench. With mHC, scores rose by 4.42 and 13.34 points on GPQA Diamond and TeamBench.

The model is now available on GitHub and Hugging Face. The team will explore more ways to improve its speed and task performance.