SenseNova Open-Sources Loop-9B Model Matching 27B Performance via Looped Architecture
SenseNova has open-sourced the Loop-9B-Preview model, utilizing a Looped Transformer architecture to extend effective computational depth through parameter sharing. The model outperforms peers in code generation and agent tasks, validating the potential for small-parameter models to approach large-model performance via increased iterations.
Full Report
IT Home reported on November 1 that SenseTime announced the open-sourcing of its SenseNova Loop-9B-Preview model on October 30.
According to official statements, the model adopts a Looped Transformer architecture, introducing looped computation to extend the effective computational depth of the model and exploring more comprehensive information interaction mechanisms. This enhances the model's capability to handle complex tasks, allowing it to rival models with 27B parameters despite having only 9B parameters.
Built upon a 9B-scale base model, SenseNova Loop-9B-Preview retains the original decoder modules for the first and last four layers, while transforming the intermediate layers into parameter-shared looped computation modules to expand effective computational depth through multiple iterations. Furthermore, the model explores information transmission mechanisms within looped computation by attempting to introduce mHC (Manifold-Constrained Hyper-Connections). This utilizes multi-path residual connections and constrained information mixing to enhance feature interaction and maintain signal propagation stability.
The report indicates that under current evaluation settings, the model outperforms other models across various domains, including code generation, office scenarios, and agentic tasks. It further surpasses Qwen3.5-27B in benchmarks such as WorkBuddy Bench, RaccoonPPT Bench, and AutomationBench. These results preliminarily demonstrate the potential of enhancing the complex task handling capabilities of small-parameter models by extending their effective computational depth.
SenseTime also evaluated the actual benefits of mHC through ablation studies. The adoption of this mechanism resulted in significant performance improvements, with scores increasing by 4.42 points on GPQA Diamond and 13.34 points on TeamBench, showcasing its application potential in complex reasoning tasks and multi-step Agent tasks. Future work will continue to explore mechanisms such as DSA and adaptive iteration counts to improve the model's inference efficiency and task performance.
Open-source links provided by IT Home:
- GitHub: https://github.com/OpenSenseNova/SenseNova-Loop
- Hugging Face: https://huggingface.co/sensenova/SenseNova-Loop-9B-Preview
Advertisement Disclaimer: External redirect links contained in this article (including but not limited to hyperlinks, QR codes, passcodes, etc.) are intended to provide additional information and save selection time. Results are for reference only. All IT Home articles containing external links include this disclaimer.