SenseTime released SenseNova Loop-9B-Preview as open source on October 30. The multimodal model has about 9 billion parameters and reuses layers for deeper processing. SenseTime says it performs at a level similar to a 27 billion parameter model in some tests.
Loop architecture
The model keeps four original decoder layers at the start and end. It runs a shared block of 24 decoder layers twice. A mechanism called Manifold-Constrained Hyper-Connections supports stable information mixing.
Training and availability
Training data includes text, images, code, and tool-use examples. The setup supports sequences up to 262,144 tokens. SenseTime released model weights, inference instructions, and training guidance under the Apache 2.0 license.
Reported performance
SenseTime reported that the model outperformed other models in code generation, office scenarios, and agent tasks. It also surpassed Qwen3.5-27B on WorkBuddy Bench, RaccoonPPT Bench, and AutomationBench. Ablation tests tied the mHC mechanism to gains of 4.42 points on GPQA Diamond and 13.34 points on TeamBench.
SenseTime plans to explore DSA and adaptive iteration counts. The project is available on GitHub and Hugging Face. These efforts aim to improve inference efficiency and task performance.