SenseTime Open-Sources SenseNova-Loop-9B-Preview with Looped Layers
SenseTime has released SenseNova-Loop-9B-Preview under the Apache 2.0 license, using a shared looped block and hyper-connections to deepen computation without increasing storage.
In short
- SenseNova-Loop-9B-Preview executes a middle block of 24 decoder layers twice, performing 56 layer passes while storing only 32 distinct layers.
- The model has approximately 9 billion total weights and roughly 14.6 billion inference parameters, released under the Apache 2.0 license.
- A two-stream manifold-constrained hyper-connections (mHC) design mixes residual data during the looped stages using Sinkhorn normalization.
- SenseTime reports the model matches 27-billion-parameter performance and outperforms Qwen3.5-27B on WorkBuddy Bench, RaccoonPPT Bench, and AutomationBench.
On October 30, 2026, artificial intelligence company SenseTime open-sourced SenseNova-Loop-9B-Preview on Hugging Face, releasing model weights and code under the permissive Apache 2.0 license. The multimodal release pairs a 9-billion-parameter weight footprint with a looped Transformer architecture that executes a large sequence of decoder layers twice per forward pass, delivering what SenseTime calculates as roughly 14.6 billion inference parameters.
Designed for agentic workflows, document comprehension, and software engineering, the model accepts text and image inputs to output text. By recycling its intermediate layer weights during inference, SenseTime claims the system achieves reasoning depth comparable to conventional 27-billion-parameter models while maintaining a smaller memory storage footprint.
How the Looped Architecture Executes
Standard autoregressive Transformer decoders evaluate token sequences by passing intermediate activations through each layer once from bottom to top. SenseNova-Loop-9B-Preview alters that progression by inserting an explicit loop over its middle layers, according to technical documentation on Hugging Face.
The backbone contains 32 distinct decoder layers organized into three computational stages. Layers 0 through 3 run once at the start of the pass. The model then enters a repeated block spanning layers 4 through 27—a contiguous stretch of 24 decoder layers—which executes twice in sequence using the exact same shared weights. Following this loop, layers 28 through 31 run once before passing representations to the output head.
Across a single forward pass, the model completes 56 layer executions while storing weights for only 32 distinct layers. This approach provides a deeper execution path without duplicating the repeated block's parameters.
Hyper-Connections and Hybrid Attention
SenseTime integrated two-stream manifold-constrained hyper-connections, abbreviated as mHC. According to reporting by Pandaily, mHC routes two parallel residual streams through the looped layers, using learned mappings to control how the model reads from and combines residual states. The stream-mixing operations are constrained through Sinkhorn normalization configured for 20 iterations.
The attention mechanism also departs from uniform self-attention. The decoder interleaves linear attention and full attention in repeating four-layer groups, placing three linear-attention layers ahead of each full-attention layer. Because the loop boundary at layer 4 aligns with these four-layer group boundaries, this attention cadence remains consistent across both the first and second passes through the shared block.
The export precision splits computational formats across components: the primary model backbone uses BF16 precision, whereas the mHC routing parameters require FP32 precision.
Benchmark Performance and Ablation Testing
SenseTime evaluated the model across several operational and agent benchmarks. According to data reported by Pandaily, SenseTime states that ablation tests showed mHC raised scores on GPQA Diamond by 4.42 points and on TeamBench by 13.34 points.
SenseTime reports that SenseNova-Loop-9B-Preview outperformed Qwen3.5-27B on WorkBuddy Bench, RaccoonPPT Bench, and AutomationBench. Model documentation also tracks evaluation results across BIRD-Dev, AIDABench, Terminal Bench 2.1, and T³-Banking. SenseTime notes that a full technical report providing comprehensive methodology and comparative metrics is still to come.
Documented training recipes for the model included long sequences of text, images, code, and supervised agent tool trajectories, reaching maximum training context lengths of up to 262,144 tokens.
Deployment Environment and Acceleration Methods
Deploying SenseNova-Loop-9B-Preview requires explicit compatibility checks due to its custom execution flow. Analysis from My Monday AI notes that the checked inference environment specifies Transformers 5.18.0 and PyTorch 2.11.0 running with CUDA 13.0, loading custom logic via remote execution flags.
Serving configurations have been integrated into third-party frameworks. The project repository provides deployment configurations for LightLLM via Docker and offers a branch targeting the SGLang inference engine via pull request #47281. The model card lists native support for MTP to accelerate decoding, while training remains underway for two additional speculative acceleration tools: Dspark and Dflash2. SenseTime indicated that subsequent research will examine DSA and adaptive loop iteration counts.
Questions people ask
What is the license for SenseNova-Loop-9B-Preview?
SenseNova-Loop-9B-Preview is released under the open-source Apache 2.0 license on Hugging Face and GitHub.
How does the looped layer mechanism work?
The model has 32 distinct layers. Layers 0 to 3 run once, layers 4 to 27 execute twice using shared weights, and layers 28 to 31 run once, resulting in 56 total layer passes per forward pass.
What modalities does the model support?
SenseNova-Loop-9B-Preview takes text and images as inputs and generates text outputs.
What hardware and software are recommended for inference?
The documented inference environment uses Transformers 5.18.0 and PyTorch 2.11.0 with CUDA 13.0, with serving templates available for LightLLM and SGLang.
Sources
- sensenova/SenseNova-Loop-9B-PreviewHugging Face, Oct 28, 2026, official source
- SenseTime Open-Sources SenseNova-Loop-9B-Preview, a Looped Model That Runs Its Middle Layers Twicepandaily.com, Oct 31, 2026
- SenseNova-Loop-9B-Preview Runs Its Own Layers Twice, and It Is Open SourceWren, Oct 31, 2026
- 5 SenseNova-Loop-9B-Preview Checks Before You DeployMy Monday AI, Oct 30, 2026
How this story was made: the newsroom picked it up from Hugging Face, Google News and pandaily.com, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (20 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published November 1, 2026 at 07:35 UTC
