SenseCore open-sources SenseNova Loop-9B-Preview using a Looped Transformer, claiming parity with 27B models
SenseCore announced the open-source release of SenseNova Loop-9B-Preview on October 30. The company said the model uses a Looped Transformer architecture that introduces recurrent computation to extend the effective computational depth of the model and explores fuller information interaction, matchi
SenseCore announced the open-source release of SenseNova Loop-9B-Preview on October 30. The company said the model uses a Looped Transformer architecture that introduces recurrent computation to extend the effective computational depth of the model and explores fuller information interaction, matching 27B-parameter models at a 9B scale. It keeps the first and last four layers of the original decoder and converts the middle layers into parameter-shared recurrent blocks iterated over multiple rounds, then explores information passing within recurrent computation and introduces manifold-constrained hyper-connections using multi-path residual connections and constrained information mixing to strengthen feature interaction and keep signal propagation stable. Under the current evaluation setup the model outperforms others in code generation, office scenarios and agentic tasks, and further exceeds Qwen3.5-27B on benchmarks including WorkBuddy Bench, RaccoonPPT Bench and AutomationBench. SenseCore evaluated mHC through ablation and reported gains of 4.42 points on GPQA Diamond and 13.34 points on TeamBench, and will continue exploring DSA and adaptive iteration counts. Code and weights were published on Github and HuggingFace.