SenseNova Loop-9B-Preview Operational

SenseNova Loop-9B-Preview is SenseTime's open-source 9B-parameter multimodal preview model. It retains the original first and last four decoder layers, replaces the middle layers with parameter-shared recurrent computation modules, expands effective computational depth through multiple iterations, and introduces the mHC mechanism to improve feature interaction stability. According to official evaluations, the model outperforms Qwen3.5-27B on office and agent tasks, delivering large-model-level capabilities on complex tasks with a small parameter count. Its weights are publicly available on GitHub and Hugging Face.

Code generation: Achieves leading performance among models in its class on code-writing tasks, supporting development applications.
Office content generation: Integrated with the Xiaohuanxiong client, it can generate office deliverables such as PPT designs and long documents.
Agent tasks: Performs strongly on multi-step agent evaluations such as AutomationBench and can complete complex, long-horizon tasks involving narratives and interaction.
Long-context processing: Officially positioned as a recurrent multimodal model for long-context and agent workflows.
Open-source commercial exploration: The weights and inference code are publicly available on GitHub and Hugging Face for community research and everyday applications.

Follow on WeChat and reply “开源” to join the AI open-source project community group
Looped Transformer architecture: Built at a 9B scale, the model retains the original first and last four decoder modules and replaces the middle layers with parameter-shared recurrent computation modules. The same set of weights is reused across multiple iterations, expanding effective computational depth without increasing parameter size and giving the small model reasoning capabilities similar to those of a deeper network.
mHC feature interaction mechanism: To improve information transfer across recurrent iterations, the model introduces Manifold-Constrained Hyper-Connections (mHC). It mixes features through multi-path residual dual streams and doubly stochastic constrained mixing matrices, enhancing inter-layer feature interaction while constraining the information manifold to maintain stable signal propagation. Ablation experiments show that this mechanism improves GPQA Diamond by 4.42 points and TeamBench by 13.34 points.
Parameter sharing and efficiency design: Shared weights in the middle recurrent layers allow the model to deliver computational depth far beyond that of a dense 9B model with approximately 9B parameters. Combined with technologies such as TBK to control per-iteration computation costs, the design aims to approach the performance of 27B-scale models on complex tasks while keeping inference costs manageable.

Environment setup: Install a version of PyTorch compatible with CUDA, then run pip install "transformers==5.18.0" accelerate safetensors.
Obtain the model: Download the sensenova/SenseNova-Loop-9B-Preview weights from Hugging Face to a local directory, ensuring it contains the configuration, tokenizer, chat template, and custom model code.
Load the model: In code, load the model with AutoModelForCausalLM.from_pretrained using trust_remote_code=True and dtype=torch.bfloat16, then call .eval().
Construct inputs: Use tokenizer.apply_chat_template to assemble the conversation prompt. Set enable_thinking=True when deep reasoning is needed.
Generate inference output: Call model.generate to generate text. use_cache is supported for acceleration, and MTP acceleration is available.
Deploy as a service: In production, use LightLLM or SGLang to provide an external API.

High parameter efficiency: With approximately 9B parameters, shared-weight recurrence can simulate 56 layers of execution depth, approaching the performance of 27B-scale models at lower cost.
Architectural innovation: It combines Looped Transformer with dual-stream mHC, using constrained multi-path residual mixing to enhance information interaction and stabilize signal propagation. It is among the first practically usable open-source implementations of its 9B class in China.
Long-context capability: Training supports a maximum context length of 262,144 tokens, making the model suitable for document understanding and long-horizon tasks.
Multimodal input: Supports text and image inputs and is officially positioned as a multimodal model for long-context and agent workflows.
Inference acceleration support: Includes a usable MTP acceleration mechanism; speculative decoding solutions such as Dspark and Dflash2 are under development.

GitHub repository: https://github.com/OpenSenseNova/SenseNova-Loop
Hugging Face model repository: https://huggingface.co/sensenova/SenseNova-Loop-9B-Preview

Long-document understanding and analysis: With its 256K long context, it can process entire contracts, financial reports, paper collections, and other large documents in one pass, generating structured summaries and answers.
Agent workflows: Performs strongly in multi-step tool calling, web operations, and task planning, making it suitable for building agents for automated office work and process execution.
Office content generation: When integrated with the Xiaohuanxiong client, it can directly generate PPT designs, data analysis reports, and animated presentation pages, covering everyday office deliverables.
Code development assistance: Its code generation capability is leading among models in its class and can handle code completion, script writing, and end-to-end development of small projects.
Multimodal content creation: Supports text and image inputs, making it suitable for rich-media content such as travel features with mixed text and images, marketing pages, and narrative websites.