SenseTime Open-Sources SenseNova Loop-9B-Preview, Claiming It Rivals 27B-Parameter Models
SenseTime announced on October 30 the open-sourcing of SenseNova Loop-9B-Preview. The model adopts a Looped Transformer architecture and expands effective computational depth through recurrent computation; the company claims that at only 9B parameters it can rival 27B-parameter models. The model retains the original decoder modules of 4 layers each at the beginning and end, converts the middle layers into parameter-sharing recurrent computation modules, and introduces mHC (Manifold-Constrained Hyper-Connections) to enhance feature interaction. The company claims that it surpasses Qwen3.5-27B on evaluations such as WorkBuddy Bench, RaccoonPPT Bench, and AutomationBench, and that in ablation experiments GPQA Diamond and TeamBench improved by 4.42 points and 13.34 points, respectively.
Why read this
AI score 60
SenseTime uses recurrent computation to increase the effective computational depth of a 9B model and provides comparative evaluations against 27B models, offering a basis for judging the small-parameter-model approach.
Topics
Body · AI translation
SenseTime Open-Sources SenseNova Loop-9B-Preview, Claims It Rivals 27B-Parameter Models
English text is being prepared.
English text is being prepared.
According to the company, the model adopts a Looped Transformer architecture, introducing recurrent computation to extend the model's effective computational depth and exploring a more thorough information interaction mechanism to enhance the model's ability to handle complex tasks, rivaling 27B-parameter models despite a limited parameter scale of only 9B.
SenseNova Loop-9B-Preview is built on a 9B-scale model as its base, retaining the first and last 4 original decoder layers and converting the intermediate layers into parameter-sharing recurrent computation modules, extending effective computational depth through multiple rounds of iteration. On this basis, the model further explores the information transfer mechanism in recurrent computation, attempting to introduce mHC (Manifold-Constrained Hyper-Connections), using multi-path residual connections and constrained information mixing to enhance feature interaction and maintain the stability of signal propagation.
According to the introduction, under the current evaluation settings, the model outperforms other models in multiple directions including code generation, office scenarios and agent tasks, and further surpasses Qwen3.5-27B in evaluations such as WorkBuddy Bench, RaccoonPPT Bench and AutomationBench. These results preliminarily demonstrate the potential of enhancing the complex task handling capability of small-parameter models by extending effective computational depth.
SenseTime further evaluated the actual benefits of mHC through ablation experiments. After adopting this mechanism, the model achieved significant performance improvements, with GPQA Diamond and TeamBench increasing by 4.42 points and 13.34 points respectively, demonstrating application potential in complex reasoning tasks and multi-step Agent tasks. In the future, it will continue to explore mechanisms such as DSA and adaptive iteration counts to improve the model's inference efficiency and task performance.
English text is being prepared.
English text is being prepared.