SenseTime open-sources 9B SenseNova-Loop-9B-Preview, a looped model, under Apache 2.0

SenseTime open-sourced SenseNova-Loop-9B-Preview on October 30, a multimodal model built on a looped Transformer design. The company says it can match models of around 27 billion parameters with only 9 billion, according to IT Home, as reported by Pandaily. Weights are on Hugging Face under the Apache 2.0 license, and code is on GitHub.

Architecture

According to the model card, the model has 32 distinct decoder layers. The first four and last four layers run once. Layers 4 through 27, a block of 24, are executed twice in sequence with shared weights. That produces 56 decoder-layer executions per forward pass while storing only 32 sets of layer weights.

The card lists about 9 billion total parameters and about 14.6 billion inference parameters.

The decoder interleaves linear and full attention. Each group of four layers has three linear-attention layers followed by one full-attention layer, and the loop boundaries line up with those groups.

The model takes text and images as input and outputs text. Its training recipes use sequences of up to 262,144 tokens, drawing on text, images, code and tool trajectories.

Manifold-constrained hyper-connections

Inside the looped section, SenseTime adds two-stream manifold-constrained hyper-connections (mHC). They carry two residual streams and use learned mappings, constrained by Sinkhorn normalization, to mix information between the streams. The company says its ablation tests showed mHC raised scores on GPQA Diamond by 4.42 points and on TeamBench by 13.34 points.

Benchmarks

SenseTime says that in its evaluation settings the model beats other models across coding, office and agent tasks. It also says the model outperforms Qwen3.5-27B on WorkBuddy Bench, RaccoonPPT Bench and AutomationBench. The benchmark chart on the model card also covers BIRD-Dev, AIDABench, Terminal Bench 2.1 and T3-Banking.

These are the company's own results. The technical report is listed as still to come.

Inference and roadmap

The model supports MTP for faster inference. Two other acceleration methods, Dspark and Dflash2, are still in training. SenseTime says it will next explore DSA and adaptive iteration counts to improve efficiency and task performance.

Pricing and hosted API availability are not yet disclosed. The release is distributed as open weights.

What this means

The design separates stored parameters from compute. The model card lists about 9 billion total parameters and about 14.6 billion inference parameters, because the middle block reuses the same weights on a second pass. Memory footprint tracks the stored weights, while per-token compute reflects the repeated layers. That is the trade the 27-billion-parameter comparison rests on.

That comparison is SenseTime's claim, based on its own evaluation settings. The listed benchmarks include several newer or less common suites, such as WorkBuddy Bench, RaccoonPPT Bench and AutomationBench. The technical report, which is still to come, is where the evaluation setup would be documented.

The Apache 2.0 license and public weights let others test the looped design and the mHC component directly. The adaptive iteration counts SenseTime plans to explore would address the fixed two-pass loop in this preview.