# 商汤发布并开源 SenseNova U1.5 技术报告，8B-MoT 原生统一模型

- 来源：SenseTime (@SenseTime_AI)
- 发布时间：2026-09-17 22:49
- AIHOT 分数：56
- AIHOT 链接：https://aihot.news/items/cmu5nx0tv0dbdroqofnjmr2wu
- 原文链接：https://x.com/SenseTime_AI/status/2100598129494294765

## AI 摘要

商汤发布 SenseNova U1.5 技术报告，推出开源 8B-MoT 原生统一模型，通过共享注意力连接理解与生成。模型采用 Pixel Shuffle 加 3×3 空间卷积实现原生 4K 生成，WISE 从 0.70 提升至 0.81，VBVR-Pro-Bench 达 68.2%，高于 Nano-Banana-Pro（56.4%）和 GPT-Image-2（50.7%）。

## 正文

🔥 𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮 𝗨𝟭.𝟱 𝗧𝗲𝗰𝗵𝗻𝗶𝗰𝗮𝗹 𝗥𝗲𝗽𝗼𝗿𝘁 𝗶𝘀 𝗼𝘂𝘁.
From capability coexistence to 𝗴𝗲𝗻𝘂𝗶𝗻𝗲 𝗰𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗶𝗻𝘁𝗲𝗿𝗮𝗰𝘁𝗶𝗼𝗻.
An open-source 8B-MoT native unified model that goes beyond "seeing + drawing."

Inside:
✨ 𝗡𝗮𝘁𝗶𝘃𝗲 𝗠𝗼𝗧 — understanding & generation connected through shared attention
✨ 𝗦𝗽𝗮𝘁𝗶𝗮𝗹𝗹𝘆 𝗝𝗼𝗶𝗻𝘁 𝗥𝗲𝗰𝗼𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝗶𝗼𝗻 — Pixel Shuffle + 3×3 spatial convolutions for native 4K generation and stronger spatial coherence
✨ 𝗦𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘇𝗲 ➡️ 𝗨𝗻𝗶𝗳𝘆 — aesthetics, OCR, infographics & editing experts consolidated via Multi-Expert On-Policy Distillation
✨ 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 ➡️ 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 — WISE 0.70 ➔ 0.81 with CoT
✨ 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 ➡️ 𝗨𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱𝗶𝗻𝗴 — RealUnify-GEU 47.5 ➔ 56.3 vs. U1-SFT
✨ 𝗩𝗶𝘀𝘂𝗮𝗹 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 — 68.2% on VBVR-Pro-Bench, ahead of Nano-Banana-Pro (56.4%) and GPT-Image-2 (50.7%)

Full training recipes — SFT, RL & multi-expert on-policy distillation — are open-sourced.

📄 Full Technical Report: https://huggingface.co/papers/2609.11929
🤗 HF: https://huggingface.co/collections/sensenova/sensenova-u15
💻 GitHub: https://github.com/OpenSenseNova/SenseNova-U1
✏️ SenseNova Studio: https://unify.light-ai.top/
⚡️Token Plan: https://www.sensenova.ai/token-plan
👾 Discord: https://discord.com/invite/BuTXPHmQub