China's tiny 9B AI just outplayed a model three times its size

In the AI race size usually wins: more parameters, more GPUs, more money. China's SenseTime went the other way and open-sourced SenseNova-Loop-9B-Preview, a 9 billion parameter model that beats Alibaba's Qwen3.5-27B, a rival three times its size, on several of the company's own benchmarks.

The trick is that part of the model runs twice. Out of 32 layers, a block of 24 is looped a second time with the same weights, so one forward pass does 56 layer computations while only 32 layers are stored. To keep information from getting lost on the second lap, SenseTime added its own scheme for mixing signals between layers, called mHC. The model reads text and images, handles long context up to 256 thousand tokens and is built for agents: documents, code and calling external tools.

So far the numbers come only from SenseTime. On the WorkBuddy Bench office test it scores 66.89 against 60.53 for Qwen3.5-27B, and on AutomationBench 13.72 against 10.17. On Terminal Bench 2.1 it still trails the bigger Qwen models, so it is not magic everywhere. A technical report is promised later.

The clip above and the sample decks on the model page were made by the Raccoon Work agent running on this model: animated websites, slides, a lunar mission simulation. The weights are on Hugging Face under the Apache 2.0 license, free to download and use even commercially, and it runs through Transformers, SGLang or LightLLM.

Verdict: small models are no longer toys. If independent tests back these numbers, a 9B model on a home GPU will start doing work that needed a rented server yesterday.