SenseTime Launches Open-Source SenseNova-Loop-9B-Preview Model

SenseTime announced the launch of the SenseNova-Loop-9B-Preview model on October 30, a multimodal model based on a looped Transformer design. The model is said to match models with around 27 billion parameters using only 9 billion, with weights available on Hugging Face under the Apache 2.0 license. The model features 32 distinct decoder layers, with the first and last four running once, while layers 4 through 27 are executed twice in sequence with shared weights. It shows superior performance in coding and office tasks, outperforming Qwen3.5-27B in several benchmarks.