SenseTime released the open-source model SenseNova Loop-9B-Preview on October 30. The model utilizes the Looped Transformer architecture, which incorporates looped computation to increase the effective depth of the model.
By retaining the first and last four decoder layers and converting the middle layers into parameter-sharing modules, the 9B model achieves performance comparable to 27B parameter models. The architecture introduces mHC, or Manifold-Constrained Hyper-Connections, to improve feature interaction and signal stability.
In testing, the model outperformed Qwen3.5-27B in benchmarks including WorkBuddy Bench, RaccoonPPT Bench, and AutomationBench. It demonstrated particular strengths in code generation, office automation, and agent-based tasks.
Ablation studies confirmed the effectiveness of the mHC mechanism. The model saw performance gains of 4.42 points in GPQA Diamond and 13.34 points in TeamBench. SenseTime plans to continue research into DSA and adaptive iteration counts to further optimize inference efficiency.
