On November 17, SenseTime (HK: 00020) officially launched SenseNova Loop-9B, a multimodal large model. It has about 9 billion parameters, accepts both text and image inputs, produces text output, and supports a maximum context length of 512K tokens. The model weights and code are open-sourced under the Apache 2.0 license.

The approach is straightforward: without adding parameters, the model runs the same set of intermediate layers twice.

SenseTime had already open-sourced a preview model, SenseNova Loop-9B-Preview, on October 30. Its weights were uploaded to GitHub and Hugging Face under the same Apache 2.0 license.

Running the Middle 24 Layers Twice

The foundation is SenseTime's in-house Looped Transformer architecture. Within a physical parameter budget of 9 billion, the model retains four conventional decoder layers at both the beginning and end, while converting the 24 middle layers, from layers 4 through 27, into a parameter-sharing looped computation module that runs twice in succession.

The second pass adds no model weights, but increases the effective computational depth from 32 layers to 56.

The model uses a hybrid attention architecture organized into four-layer groups, each comprising three linear-attention layers and one full-attention layer. One detail matters: the loop boundaries align with complete four-layer groups, keeping the attention pattern consistent across both passes. If a loop were to split a four-layer group, the two passes would encounter different attention structures, undermining the shared-layer design.

On top of this is an mHC flow-constrained hyperconnection mechanism. It uses multiple residual connections and a flow-mixing matrix constrained by Sinkhorn normalization to strengthen feature interaction between layers and maintain stable signal propagation. Ablation tests show that the mechanism improved scores on GPQA Diamond and TeamBench by 4.42 and 13.34 points, respectively.

The Trade-Off: Computing Each Token Twice

The parameter count is unchanged, but the number of layer executions rises from 32 to 56. Given the claims of “no additional model weights” and “56 layer computations,” the computation required for each token goes up, not down.

The savings come elsewhere: model weight size and GPU memory usage. SenseTime says the model has “extremely low GPU memory usage” and can be “deployed on a single midrange data center GPU.”

That distinction defines the value of this approach. It addresses whether a model fits, not whether it runs fast. For companies that need to deploy a model in their own data centers or close to production systems, capacity is the real barrier. For cloud providers spreading costs across API calls at scale, speed matters more. The same technology could draw opposite assessments from these two customer groups.

This approach moves in the opposite direction from today's mainstream model-lightweighting techniques. Mixture-of-Experts models rely on sparse activation, invoking only a subset of parameters at a time to reduce computation per pass. Looped layers run the same parameters twice, increasing computation in exchange for greater depth. Most competing vendors use MoE for lightweight models. Both approaches seek to make small models do the work of larger ones, but one saves computing power while the other saves GPU memory, leaving little overlap between their target use cases.

Leading on Six of Eight Benchmarks

SenseTime published results for eight benchmarks, with Loop-9B outperforming comparison models in the same parameter class on six. It scored 66.89 on WorkBuddy Bench for office tasks, ahead of Qwen3.5-27B's 60.53, and 70.70 on RaccoonPPT Bench, versus 59.08 for the comparison model.

There is an important distinction here. SenseTime says the model “delivers complex-task performance comparable to a 27B-parameter model,” but that claim rests on the two results above: WorkBuddy and RaccoonPPT both cover office and document tasks. The separate claim that Loop-9B leads on six benchmarks against models of a similar parameter scale refers to a different set of competitors, not 27B models.

Taken together, the accurate conclusion is that Loop-9B outperformed a 27B comparison model on office and document tasks. That result cannot be directly extrapolated to every task, and SenseTime did not provide like-for-like data across the full benchmark suite.

Preview Arrived 18 Days Before the Official Release

The official release on November 17 came 18 days after the preview version was released on October 30.

SenseTime says it open-sourced the preview first to gather community feedback, then iterated on the model's capabilities before releasing the official version. This is unusual among open-source model developers: most teams either issue patches after release or launch the final version directly, rather than making an intermediate build publicly available.

For users, the practical implication is that code already working with the preview can probably switch directly to the official release. But anyone who put the preview into production during those 18 days had to bear the risk themselves. That is also why many companies prefer to wait for an official release: getting access 18 days earlier could mean spending twice as long on regression testing.

Where Enterprise Deployments Get Stuck

SenseTime identifies three common barriers to enterprise adoption of large models over the past two years: high deployment costs, steep hardware procurement requirements, and inference latency that is difficult to accommodate. Small models in the 9B class are cheaper but often cannot handle complex tasks. Yet applications such as office automation, long-document auditing, and agent tool use require deeper reasoning.

Loop-9B's answer is to deliver 27B-class task performance in a 9B-sized model, bringing deployment down to a single midrange GPU. SenseTime also says enterprise customers already connected to its AI services can upgrade at low cost without replacing their existing hardware.

Of those three pain points, however, this model clearly addresses the first: making deployment affordable and ensuring the model fits. Actual inference latency after running the looped layers twice remains to be seen, and SenseTime has not disclosed that figure. Whether it truly lowers the hardware barrier will depend on which GPUs customers already own.

SenseTime positions the model around a “small parameter count, high capability” strategy, arguing that industry competition is shifting from parameter counts and scale toward architecture and deployment efficiency. Whether that assessment holds will depend on two things: whether independent evaluations can reproduce the “27B-class” claim, and whether enterprises can accept the latency introduced by running layers twice.

At least for office and document tasks, the reported figures are concrete: 66.89 versus 60.53, and 70.70 versus 59.08. Those numbers are easier to verify than any adjective.