SenseTime Open-Sources SenseNova Loop-9B-Preview: Exploring Small Model Performance via Looped Transformer Architecture
SenseTime releases and open-sources the SenseNova Loop-9B-Preview model based on Looped Transformer architecture, claiming it matches 27B models in multiple tasks.
SenseTime officially open-sourced the SenseNova Loop-9B-Preview model from its Riri Xin series on October 31, 2026. This large language model employs an innovative Looped Transformer architecture designed to extend the effective computational depth of the model through cyclic computation, aiming to achieve or surpass the performance of traditional 27B parameter models with only 9B parameters.
The core design retains the first and last four layers of the original decoder modules while transforming the intermediate layers into parameter-sharing cyclic computation modules that increase information interaction depth through multiple iterations. To maintain signal propagation stability and enhance feature interaction, SenseTime introduced the mHC (Manifold-Constrained Hyper-Connections) mechanism, utilizing multi-path residual connections for constrained information mixing. This architecture attempts to address the issue where current LLM capability improvements rely excessively on scaling parameter counts, exploring a new path to release model potential while controlling parameter size.
According to internal evaluation results published by SenseTime, SenseNova Loop-9B-Preview demonstrates strong performance across code generation, office scenarios, and agent tasks. Specifically, the model claims to outperform Qwen3.5-27B in benchmarks such as WorkBuddy Bench, RaccoonPPT Bench, and AutomationBench. Ablation studies indicate that introducing the mHC mechanism improved scores on GPQA Diamond and TeamBench by 4.42 and 13.34 points respectively. It is important to note that these are vendor-reported results and have not yet been independently reproduced by third parties.
Currently, research community explorations of Looped Transformers remain largely academic, lacking direct open-source practices for daily use. The release of SenseNova Loop-9B-Preview fills this gap, with weights uploaded to GitHub and Hugging Face. SenseTime stated that future work will continue to explore scaling potential at larger sizes and adaptive 'fast/slow thinking' mechanisms to further improve inference efficiency.