Reportedly: SenseTime open-sources SenseNova Loop-9B-Preview, claiming performance comparable to 27B-parameter models
Reported by IT Home · not yet confirmed by the company or a second independent outlet. We update this page when it is.
SenseTime open-sourced SenseNova Loop-9B-Preview on October 30, 2026, a 9B-parameter model using a Looped Transformer architecture that the company says matches 27B-parameter models. The model outperforms Qwen3.5-27B on WorkBuddy Bench, RaccoonPPT Bench, and AutomationBench. Availability details beyond the open-source release were not specified, IT Home reported.
China context
- Original name
- 商汤科技
- Outside China
- Open weights · ithome.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can access the model weights and evaluate whether the 9B model meets their needs for code generation, office tasks, and agent workflows, potentially reducing compute costs compared to 27B models.
- For investors
- SenseTime's move into open-weights small models with competitive performance claims may affect the market for mid-sized models; investors should track independent benchmark results and developer adoption.
Translated from Chinese. Quotes and facts link to the original sources.
SenseTime announced the open-source release of SenseNova Loop-9B-Preview, a 9B-parameter model built on a Looped Transformer architecture. The design keeps the first and last four decoder layers as original modules and converts the middle layers into a parameter-shared recurrent computation module, extending effective computational depth through multiple iterations. It also introduces Manifold-Constrained Hyper-Connections (mHC) to enhance feature interaction and signal propagation stability. The company reports that the model surpasses Qwen3.5-27B on benchmarks including WorkBuddy Bench, RaccoonPPT Bench, and AutomationBench, and performs well in code generation, office scenarios, and agent tasks.
The Looped Transformer architecture reuses parameters in the middle layers to increase effective depth without increasing parameter count, a technique aimed at improving complex-task performance in small models. The mHC mechanism uses multi-path residual connections with constrained information mixing to stabilize training and inference. Benchmark results are company-reported and have not been independently verified.
Developers outside China gain access to a 9B-parameter open-weights model that SenseTime claims matches 27B-parameter performance, potentially reducing inference costs for complex tasks. Competitors offering 27B-class models may face pressure on price-performance if independent evaluations confirm the claims.
For enterprises and developers, a smaller model with claimed 27B-level performance could lower deployment and inference costs while maintaining capability for code generation, office automation, and agent tasks. The open-source license may facilitate commercial use, though specific terms were not detailed in the evidence.
Independent evaluation of SenseNova Loop-9B-Preview on the cited benchmarks and real-world tasks will determine whether the performance claims hold. Adoption by developers and integration into downstream applications are observable next signals.