Sensenova loop-9b-preview model released by SenseTime, claims to match 27b-parameter models

On November 1, 2026, IT Home reported that SenseTime announced the open-sourcing of its SenseNova Loop-9B-Preview model on October 30. The model uses a Looped Transformer architecture, incorporating cyclic computation and expanding the effective depth of model calculations. It aims to enhance information interaction mechanisms, improving performance on complex tasks with only 9B parameters, matching models with 27B parameters. The model is based on a 9B-scale foundation, retaining the first and last four layers of the original decoder modules, while transforming the middle layers into parameter-shared cyclic computation modules. It explores information transfer mechanisms within cyclic computation, introducing mHC (Manifold-Constrained Hyper-Connections) to enhance feature interaction and maintain signal stability. In current evaluations, the model outperforms others in code generation, office scenarios, and agent tasks, surpassing Qwen3.5-27B in WorkBuddy Bench, RaccoonPPT Bench, and AutomationBench. SenseTime also conducted ablation studies, showing significant performance gains with mHC, including improvements of 4.42 points on GPQA Diamond and 13.34 points on TeamBench.