您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：SenseNova Loop-9B-Preview 只是起点。
商汤开源日日新 SenseNova Loop-9B-Preview。这是商汤面向下一代模型架构的一项前沿探索：它采用创新 Looped Transformer 架构，引入循环计算、扩展模型的有效计算深度，并探索更充分的信息交互机制，提升模型在复杂任务上的处理能力，在仅 9B 的有限参数规模下，媲美 27B 参数大小模型。
随着大模型能力持续演进，模型能力提升不再只有扩大参数规模一条路径。如何在控制参数规模的同时，通过增加有效计算深度进一步释放模型潜力，正成为值得探索的方向。近期，循环计算正成为大模型架构研究的重点方向，国内外研究社区围绕 Looped Transformer 展开了攻关探索。但目前的探索多停留在学术研究层面，社区尚缺乏可直接面向日常使用的 Looped Transformer 模型。
此次开源的 SenseNova Loop-9B-Preview，为社区提供了一个 9B 级 Looped Transformer 的开源实践，既填补了这一空白，体现了商汤对下一代大模型架构的持续探索与思考。
不过，如何让多轮计算真正转化为模型能力的提升，尤其是在信息传递、特征交互和训练稳定性等方面，仍有进一步探索的空间。现有相关工作仍以研究验证为主，如何在更大模型规模下兼顾计算深度、模型能力与实际应用需求，也有待持续推进。
循环结构与信息交互机制
SenseNova Loop-9B-Preview 以 9B 尺度模型为基座，保留首尾各 4 层原始解码器模块，将中间层改造为参数共享的循环计算模块，通过多轮迭代扩展有效计算深度。
在此基础上，模型进一步探索循环计算中的信息传递机制，尝试引入 mHC（Manifold-Constrained Hyper-Connections），利用多路残差连接与受约束的信息混合，增强特征交互并维持信号传播的稳定性。整体设计旨在控制参数规模的同时，探索提升模型推理能力的新路径。
SenseNova Loop-9B-Preview 在保持 9B 级参数规模的同时，展现出较强的跨任务性能。在当前评测设置下，模型在代码生成、办公场景和智能体任务等多个方向上均优于其他模型，并在 WorkBuddy Bench、RaccoonPPT Bench 和 AutomationBench 等评测中进一步超过 Qwen3.5-27B。这些结果初步展现了通过扩展有效计算深度，提升小参数模型复杂任务处理能力的潜力。
商汤进一步通过消融实验评估了 mHC 的实际收益。采用该机制后，模型取得明显的性能提升，其中 GPQA Diamond 和 TeamBench 分别提高了 4.42 分和 13.34 分，展现出在复杂推理任务以及多步骤 Agent 任务中的应用潜力。后续还将继续探索 DSA、自适应迭代次数等机制，以提升模型的推理效率与任务表现。
SenseNova Loop-9B-Preview 只是起点。面向下一代 SenseNova，商汤将进一步探索 Looped Transformer 在更大模型规模下的 Scaling 潜力、自适应“快慢思考”机制，深度挖掘参数共享的更多潜能，以更灵活的计算方式、更高的参数效率，持续突破模型能力与推理效率的边界。
模型权重开源：
Github：https://github.com/OpenSenseNova/SenseNova-Loop
Hugging Face：https://huggingface.co/sensenova/SenseNova-Loop-9B-Preview
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: SenseNova Loop-9B-Preview is just the starting point.
It's the new SenseNoova Loop-9B-Preview. This is a front-line exploration for the next generation of model structures: using an innovative Looped Transformer structure, it introduces the effective computation depth of the cycle computing, extension models, and explores a fuller information interaction mechanism to enhance the model ' s processing capacity on complex missions, comparable to the 27B parameter size model at the limited scale of only 9B.
As LLM's capabilities continue to evolve, model capacity is no longer only one path to increase the size of the parameters. The potential for further release of models by increasing the effective computation depth while controlling the size of parameters is becoming a promising direction. In the recent past, circulatory calculations are becoming the focus of LLM structural research, and national and international research communities have been engaged in critical research around Looped Transformer. However, much of the current exploration is at the academic research level, and the community lacks a Looped Transformer model that can be used directly for daily use.
This open-source SenseNoova Loop-9B-Preview provides a 9B-class Looped Transformer open-source practice that fills this gap and reflects the ongoing exploration and reflection of the next generation of LLM structures.
However, there is still room for further exploration of how multi-wheel computing can be truly translated into enhanced model capabilities, particularly in the areas of information transfer, characterization interaction and training stability. Existing work is still based on research validation, and the balance between computational depth, modelling capacity and practical application needs at a more LLM scale will need to be continued.
Cycle structure and information interface
SenseNoova Loop-9B-Preview uses a 9B scale model as its base, preserves the 4-story original decoder module at the end, transforms the middle layer into a recycle calculation module for the sharing of parameters, and expands the Depth effectively through multiple rotations.
On this basis, the model further explores the information transfer mechanisms in the cycle calculations, attempts to introduce mHCs (Manifold-Constrained Hyper-Connects), uses multiple-road disability connections to mix bound information, enhances feature interaction and maintains the stability of signal transmission. The overall design is designed to control the size of the parameters while exploring new pathways to enhance model reasoning capabilities.
SenseNoova Loop-9B-Preview displays a strong cross-mission performance while maintaining the size of the 9B level parameter. Under the current evaluation settings, models are better than other models in many directions, including code generation, office scenes and AI agent missions, and further exceed Qwen3.5-27B in the WorldBudy Bench, RaccoonPPT Bench and Automation Bench assessments. These results provide an initial demonstration of the potential for enhancing the complex mission processing capacity of small-parameter models by expanding effective computation depth.
Comfort further assessed the actual benefits of mHC through the digestion experiment. Following the introduction of the mechanism, the model achieved significant performance enhancement, with GPQA Diamond and TeamBench increasing by 4.42 and 13.34, respectively, to demonstrate the potential for application in complex reasoning tasks and multi-step Agent missions. Mechanisms such as DSA, the number of adaptations will continue to be explored to enhance the reasoning efficiency and mission performance of the model.
SenseNoova Loop-9B-Preview is just the beginning. For the next generation, Sensenova, Comfort will explore Loped Transformer's potential at a more LLM scale, adapting itself to a “fast-think” mechanism, deepening the potential of the more shared parameters, in a more flexible way of calculating, more efficient parameters, and continuing to breach the boundaries of modeling and reasoning.
Model weight reopen source:
Github：https://github.com/OpenSenseNova/SenseNova-Loop
Hugging Face：https://huggingface.co/sensenova/SenseNova-Loop-9B-Preview
For your account security, please verify the mailbox.
Please verify your mailbox.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.