From the source
SenseTime has open-sourced SenseNova Loop-9B-Preview, a 9B-parameter model built on a Looped Transformer architecture that the company says matches 27B-scale models by expanding effective compute depth rather than parameter count.
- The company frames the release as filling a gap: looped-computation research has stayed largely academic, and the community lacked a Looped Transformer model meant for everyday use.
- The model keeps four original decoder layers at each end and turns the middle layers into a parameter-sharing looped module that iterates to extend effective compute depth.
- It also introduces mHC (Manifold-Constrained Hyper-Connections), using multiple residual connections and constrained information mixing to strengthen feature interaction and keep signal propagation stable.
- SenseTime says the model beats other models on code generation, office scenarios and agent tasks, and surpasses Qwen3.5-27B on WorkBuddy Bench, RaccoonPPT Bench and AutomationBench.
- Ablation tests credited mHC with gains of 4.42 points on GPQA Diamond and 13.34 points on TeamBench.
- Weights are on
GitHub and
Hugging Face, and the model is used in SenseTime's Xiaohuanxiong client.





