Image: SenseTime

SenseTime has released SenseNova-Loop-9B-Preview, an open source AI model that tries to get more reasoning out of a small model by running part of it twice. The company announced the release on X and published the weights on Hugging Face with code and documentation on GitHub, under the Apache 2.0 license.

The model accepts text and images and produces text. SenseTime describes it as built for long context and agentic workflows, meaning tasks where a model works with documents, writes code and calls external tools over several steps.

The central idea is what SenseTime calls scaling compute depth rather than model size. Most language models pass input through each layer once. In this model, according to the model card, a block of 24 of its 32 decoder layers is applied twice in a row with the same weights. That gives 56 layer computations on every forward pass while only 32 sets of layer weights are stored. The model card lists about 9 billion total parameters and about 14.6 billion inference parameters.

To help information move through the repeated block, SenseTime uses a technique it calls manifold-constrained hyper-connections, or mHC. The model card says it carries two residual streams around the looped layers, with learned mappings deciding how each layer reads from and mixes those streams. The decoder also mixes attention types, with three linear attention layers followed by one full attention layer in each group of four. Training recipes use sequences of up to 262,144 tokens.

Image: SenseTime

On performance, SenseTime claims the 9B model delivers 27B-class results on complex tasks and outperforms Alibaba's Qwen3.5-27B on WorkBuddy Bench, RaccoonPPT Bench and AutomationBench. In the company's published chart, the model scores 66.89 on WorkBuddy Bench against 60.53 for Qwen3.5-27B, 70.70 on RaccoonPPTBench against 59.08, and 13.72 on AutomationBench against 10.17. SenseTime also says mHC alone raises scores by 4.42 points on GPQA Diamond and 13.34 points on TeamBench. These are the company's own numbers and had not been independently verified at the time of this writing.

The same chart shows the model does not lead everywhere. On Terminal Bench 2.1, SenseTime reports 40.45, behind Qwen3.5-27B at 47.19, and on the banking benchmark it scores 18.56 against 19.59 for Qwen3.5-27B. A footnote in the chart describes RaccoonPPTBench as an internal benchmark, and the showcase demos on the model page, including animated HTML pages, slide decks and data analysis, are credited to SenseTime's own Raccoon Work product.

Some pieces are still missing. The model card lists the technical report as to be announced. It says multi-token prediction is available for faster inference, while two other speedup methods, Dspark and Dflash2, are still in training. The documented setup uses Transformers 5.18.0 with custom model code, and SenseTime gives deployment examples for its LightLLM serving framework.

The release follows SenseTime's earlier work on looping in image generation. In September, its OpenSenseNova team published Looped-DiT, a 260 million parameter image model that the team said beat a model 6.5 times larger. In its announcement, SenseTime said the next steps for the Loop line are larger models, adaptive fast and slow thinking, and more efficient reasoning.

Introducing SenseNova Loop-9B-Preview, now open source! Scaling compute depth, not just model size: our 9B Looped Transformer uses parameter-shared loops + mHC (Manifold-Constrained Hyper-Connections) to improve information flow and reasoning. pic.twitter.com/HOAEm8VkMn

— SenseTime (@SenseTime_AI) October 30, 2026