跳到正文
SenseTime· @SenseTime_AI · X·· 2 小时前AI 评分54
AI 导读

商汤发布并开源 SenseNova Loop-9B-Preview,采用参数共享 Looped Transformer 加 mHC(Manifold-Constrained Hyper-Connections)提升信息流与推理能力。

正文 · 原文

𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮 𝗟𝗼𝗼𝗽-𝟵𝗕-𝗣𝗿𝗲𝘃𝗶𝗲𝘄, 𝗻𝗼𝘄 𝗼𝗽𝗲𝗻 𝘀𝗼𝘂𝗿𝗰𝗲!

𝗦𝗰𝗮𝗹𝗶𝗻𝗴 𝗰𝗼𝗺𝗽𝘂𝘁𝗲 𝗱𝗲𝗽𝘁𝗵, 𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝗺𝗼𝗱𝗲𝗹 𝘀𝗶𝘇𝗲: our 9B Looped Transformer uses parameter-shared loops + mHC (Manifold-Constrained Hyper-Connections) to improve information flow and reasoning.

📊 𝗛𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀:
🔹Delivers 𝟮𝟳𝗕-𝗰𝗹𝗮𝘀𝘀 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 on complex tasks, outperforming Qwen3.5-27B on 𝗪𝗼𝗿𝗸𝗕𝘂𝗱𝗱𝘆 𝗕𝗲𝗻𝗰𝗵, 𝗥𝗮𝗰𝗰𝗼𝗼𝗻𝗣𝗣𝗧 𝗕𝗲𝗻𝗰𝗵 & 𝗔𝘂𝘁𝗼𝗺𝗮𝘁𝗶𝗼𝗻𝗕𝗲𝗻𝗰𝗵
🔹𝗺𝗛𝗖 𝗯𝗼𝗼𝘀𝘁𝘀 𝘀𝗰𝗼𝗿𝗲𝘀 by 𝟰.𝟰𝟮 𝗽𝗼𝗶𝗻𝘁𝘀 on GPQA Diamond and 𝟭𝟯.𝟯𝟰 on TeamBench
🔹Strong across 𝗰𝗼𝗱𝗶𝗻𝗴, 𝗼𝗳𝗳𝗶𝗰𝗲 & 𝗮𝗴𝗲𝗻𝘁 𝘁𝗮𝘀𝗸𝘀, with 𝗶𝗻𝘁𝗲𝗿𝗮𝗰𝘁𝗶𝘃𝗲, 𝗮𝗻𝗶𝗺𝗮𝘁𝗲𝗱 𝗛𝗧𝗠𝗟 𝗱𝗲𝗺𝗼𝘀

🔬 𝗡𝗲𝘅𝘁: Scaling to larger models, adaptive fast/slow thinking & more efficient reasoning.

🤗 Model: https://huggingface.co/sensenova/SenseNova-6.8-Loop-9B-Preview
💻 GitHub: https://github.com/OpenSenseNova/SenseNova-Loop
👾Discord: https://discord.com/invite/BuTXPHmQub

来源:SenseTime · x.com