SenseTime releases SenseNova-Loop-9B-Preview with shared decoder layers

SenseTime released SenseNova-Loop-9B-Preview on October 30, as detailed on the model's Hugging Face page. The multimodal model accepts text and images, and generates text; its documented training recipes use sequences of up to 262,144 tokens.

The model contains about 9 billion total parameters and 32 distinct decoder layers. It applies a shared block of 24 layers twice, resulting in 56 decoder layer executions per forward pass while storing 32 sets of layer weights. A two stream connection system manages information flow through the repeated section.

SenseNova-Loop-9B-Preview uses a repeating attention pattern with three linear attention layers followed by one full attention layer. Its weights are available under the Apache 2.0 license, and documented deployment options include LightLLM and development versions of SGLang and vLLM.

SenseTime says the design lets the 9 billion parameter model match models of around 27 billion parameters and that it outperforms Qwen3.5-27B on its WorkBuddy, RaccoonPPT and AutomationBench tests, Pandaily reports, citing IT Home. These are the company's own results, and the technical report has not been published yet.