The article title: 寒武纪 completes Day 0 adaptation of SenseTime SenseNova Loop-9B, opening up the inference pipeline for a new recurrent computation architecture
Body: IT Home, November 18 news: SenseTime Technology officially released the SenseNova Loop-9B model last night.

寒武纪 officially announced that it has completed Day 0 adaptation of SenseTime's newly released model, SenseNova Loop-9B.
According to the introduction, the model reuses the parameters of the model layers through a Loop structure, and combined with manifold-constrained hyper-connections (mHC), expands computational depth under a limited parameter scale. It targets complex reasoning, code, tool calling, and multimodal tasks, further optimizing the model's performance in Agent, code, and data analysis tasks, and continuously exploring the capability boundaries of recurrent computation.
寒武纪 stated that, in response to the architectural characteristics of SenseNova Loop-9B, 寒武纪 completed Day 0 adaptation based on vllm_mlu, opening up the model's inference pipeline on the MLU platform. This adaptation mainly focuses on three directions: the Loop recurrent structure, MLU operator adaptation, and mHC routing optimization. IT Home attaches the specific adaptation contents of the three directions as follows:
1. Loop recurrent layer unfolding: statically unfolded computation, reusing the standard inference mechanism
The Loop recurrent structure is one of the core technical points in the adaptation of the SenseNova Loop-9B model. SenseNova Loop-9B uses about 9B model parameters and, by recurrently executing shared intermediate layers, expands 32 physical decoder layers into 56 layer computations. Thanks to the efficient Loop recurrent structure, the model's official 4bit quantized weights are only about 4.9GB, and the deployment VRAM requirement is as low as within 10 GB, which can significantly broaden the range of deployable hardware and help lower the threshold for local and private deployment. To be compatible with vLLM's existing model execution and scheduling mechanisms, 寒武纪 adopted a static unfolding scheme, converting the recurrent computation into 56 sequentially executed layers, enabling the model to reuse vLLM's standard scheduling and KV cache mechanisms. During implementation, the recurrent layers at different execution slots share the same weight storage, and the VRAM usage is the same as that of a 32-layer model.
The Loop recurrent structure is one of the core technical points in the adaptation of the SenseNova Loop-9B model. SenseNova Loop-9B uses about 9B model parameters and, by recurrently executing shared intermediate layers, expands 32 physical decoder layers into 56 layer computations.
Thanks to the efficient Loop recurrent structure, the model's official 4bit quantized weights are only about 4.9GB, and the deployment VRAM requirement is as low as within 10 GB, which can significantly broaden the range of deployable hardware and help lower the threshold for local and private deployment.
To be compatible with vLLM's existing model execution and scheduling mechanisms, 寒武纪 adopted a static unfolding scheme, converting the recurrent computation into 56 sequentially executed layers, enabling the model to reuse vLLM's standard scheduling and KV cache mechanisms. During implementation, the recurrent layers at different execution slots share the same weight storage, and the VRAM usage is the same as that of a 32-layer model.
2. MLU operator adaptation: reusing mature operator implementations, supporting multimodal inference
For core computation modules in the model such as Attention, Linear Attention, and MLP, 寒武纪 replaced them with the existing Qwen3-Next series MLU operator implementations in vllm_mlu, reusing existing operator capabilities to quickly achieve model adaptation. In the multimodal part, it reuses Qwen3.5-VL's MLUQwen3_VisionTransformer and multimodal Processor to implement image input processing and visual feature extraction, opening up the inference pipeline in text and image input scenarios.
For core computation modules in the model such as Attention, Linear Attention, and MLP, 寒武纪 replaced them with the existing Qwen3-Next series MLU operator implementations in vllm_mlu, reusing existing operator capabilities to quickly achieve model adaptation.
In the multimodal part, it reuses Qwen3.5-VL's MLUQwen3_VisionTransformer and multimodal Processor to implement image input processing and visual feature extraction, opening up the inference pipeline in text and image input scenarios.
3. mHC routing performance optimization: fusing normalization computation, reducing Kernel Launch overhead
mHC is an important component of the model's recurrent computation. Its dual-stream information routing involves normalization and Sinkhorn iterative computation, which places requirements on operator execution efficiency. For mHC routing computation, 寒武纪 carried out special optimization: Normalization optimization: implement normalization computation through the TMO fusion interface to reduce operator execution overhead. Sinkhorn optimization: use CNAgent Triton Skill to automatically generate a Triton fused operator, integrating the Sinkhorn computation into a single Triton Kernel, reducing the scheduling overhead caused by multiple Kernel Launches.
mHC is an important component of the model's recurrent computation. Its dual-stream information routing involves normalization and Sinkhorn iterative computation, which places requirements on operator execution efficiency.
For mHC routing computation, 寒武纪 carried out special optimization:
- Normalization optimization: implement normalization computation through the TMO fusion interface to reduce operator execution overhead.
- Sinkhorn optimization: use CNAgent Triton Skill to automatically generate a Triton fused operator, integrating the Sinkhorn computation into a single Triton Kernel, reducing the scheduling overhead caused by multiple Kernel Launches.
- "SenseTime releases the SenseNova Loop-9B model: deployment VRAM requirement as low as within 10 GB"