ModelsSelected

Cambricon completes Day 0 adaptation of SenseNova Loop-9B

寒武纪 Day 0 完成商汤 SenseNova Loop-9B 适配,打通新型循环计算架构推理链路

Why we picked it

Day 0 hardware co-optimisation is becoming the standard play for Chinese AI stacks, and this pairing shows how recurrent-depth architectures like Loop-9B trade parameter count for compute depth to run in under 10GB of memory. Teams deploying private inference on domestic MLU accelerators can evaluate it immediately via vllm_mlu, comparing it against dense 9B-class models on memory footprint and reasoning quality. The reuse of Qwen3-Next and Qwen3.5-VL operators also hints at growing kernel-level convergence across Chinese model families.

On 17 Nov 2025 SenseTime released its SenseNova Loop-9B model, and Cambricon (寒武纪) announced Day 0 adaptation of the model on its MLU hardware platform, completing the inference pipeline based on vllm_mlu. The adaptation covered three areas: expanding the Loop recurrent layers, MLU operator support, and optimising the mHC routing computation.

Body

SenseNova Loop-9B uses about 9B parameters and extends computation depth by reusing parameters through a Loop structure with manifold-constrained hyper-connections (mHC), unrolling 32 physical decoding layers into 56 layer computations. It targets complex reasoning, code, tool calling and multimodal tasks. Cambricon statically unrolled the loop into 56 sequentially executed layers so the model reuses vLLM's standard scheduling and KV cache, with loop layers sharing one weight store so memory usage matches a 32-layer model. Official 4-bit quantised weights are about 4.9GB, allowing deployment under 10GB of memory. For operators, Cambricon reused existing Qwen3-Next MLU kernels for Attention, Linear Attention and MLP, and Qwen3.5-VL's vision transformer for multimodal input; mHC routing received fused normalisation to cut kernel launch overhead.

Day 0 support on domestic accelerators lowers the barrier for local and private deployment of recurrent-compute models, giving enterprises running on MLU hardware immediate access to SenseTime's newest reasoning model without waiting for community ports.

Original · IT之家

寒武纪 Day 0 完成商汤 SenseNova Loop-9B 适配,打通新型循环计算架构推理链路

IT之家 11 月 18 日消息, 商汤科技昨日晚间正式发布 SenseNova Loop-9B 模型 。 寒武纪官方宣布, 已完成对商汤新发布模型日日新 SenseNova Loop-9B 的 Day 0 适配 。 据介绍,该模型通过 Loop 结构复用模型层的参数,并结合流形约束超连接(mHC),在有限的参数规模下扩展计算深度,面向复杂推理、代码、工具调用和多模态任务,进一步优化模型在 Agent、代码与数据分析任务中的表现,持续探索循环计算的能力边界。 寒武纪表示,针对 SenseNova Loop-9B 的架构特点, 寒武纪基于 vllm_mlu 完成 Day 0 适配,打通模型在 MLU 平台上的推理链路 。本次适配主要围绕 Loop 循环结构、MLU 算子适配、mHC 路由优化三个方向展开。IT之家附三大方向具体适配内容如下: 1. Loop 循环层展开:静态展开计算,复用标准推理机制 Loop 循环结构是 SenseNova Loop-9B 模型适配的核心技术点之一。SenseNova Loop-9B 使用约 9B 的模型参数,通过循环执行共享的中间层,将 32 个物理解码层扩展为 56 次层计算。 正是得益于 Loop 的高效循环结构,模型官方 4bit 量化后权重仅约 4.9GB,部署显存需求低至 10 GB 内,能显著拓宽可部署的硬件范围,有助于降低本地与私有化部署门槛。 为了兼容 vLLM 现有的模型执行与调度机制,寒武纪采用静态展开方案,将循环计算转换为 56 个顺序执行的层,使模型能够复用 vLLM 标准调度和 KV cache 机制。在实现过程中,不同执行槽位的循环层共享同一份权重存储,显存占用与 32 层模型相同。 2. MLU 算子适配:复用成熟算子实现,支持多模态推理 针对模型中的 Attention、Linear Attention 和 MLP 等核心计算模块,寒武纪将其替换为 vllm_mlu 中已有的 Qwen3-Next 系列 MLU 算子实现,复用现有算子能力,从而快速实现模型适配。 在多模态部分,复用 Qwen3.5-VL 的 MLUQwen3_VisionTransformer 和多模态 Processor,实现图像输入处理及视觉特征提取,打通文本与图像输入场景下的推理链路。 3. mHC 路由性能优化:融合归一化计算,降低 Kernel Launch 开销 mHC 是模型循环计算中的重要组成部分,其双流信息路由涉及归一化及 Sinkhorn 迭代计算,对算子执行效率提出了要求。 针对 mHC 路由计算,寒武纪进行了专项优化: 归一化优化: 通过 TMO 融合接口实现归一化计算,减少算子执行开销。 Sinkhorn 优化:使用 CNAgent Triton Skill 自动生成 Triton 融合算子,将 Sinkhorn 计算整合为单个 Triton Kernel,减少多次 Kernel Launch 带来的调度开销。 相关阅读: 《 商汤发布 SenseNova Loop-9B 模型:部署显存需求低至 10 GB 内 》

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include SenseNova Loop-9B, SenseTime, Cambricon, Qwen3-Next.

Source attribution: curated from IT之家.