AI 导读
商汤 SenseNova U1 公开完整技术报告，并开源 SenseNova-U1-A3B-MoT（38B-A3B MoE）模型权重，仅 3B 激活参数。该模型采用免 VE、免 VAE 的近无损视觉接口与原生 Mixture-of-Transformers 骨干，共同进行 AR 与像素空间流匹配训练。技术报告包含 6 阶段训练配方以及 RL 后训练与蒸馏，论文地址为 arxiv.org/abs/2605.12500。
正文
While everyone is scaling up VE/VAE, SenseNova U1 goes VE‑free + MoT.
Full tech report now public.
And SenseNova-U1-A3B-MoT (38B-A3B MoE) weights are now open-sourced.
🔥 New week, New 𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮-𝗨𝟭 Drop — and this one goes Deep!🔥 📄 𝗧𝗵𝗲 𝗳𝘂𝗹𝗹 𝗧𝗲𝗰𝗵𝗻𝗶𝗰𝗮𝗹 𝗥𝗲𝗽𝗼𝗿𝘁 𝗶𝘀 𝗢𝗨𝗧 — the most detailed disclosure yet of how to build a frontier Native Multimodal Model. Inside: ✨ Near-lossless visual interface (no VEs, no VAEs) ✨ Native Multimodal Unified Modeling ✨ Joint AR + pixel-space flow matching training ✨ Native Mixture-of-Transformers backbone ✨ 6-stage training recipe + RL post-training + distillation If you work on NMM, this is the playbook. 🤗 One more thing: 𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮-𝗨𝟭-𝗔𝟯𝗕-𝗠𝗼𝗧 (𝟯𝟴𝗕-𝗔𝟯𝗕 𝗠𝗼𝗘) 𝘄𝗲𝗶𝗴𝗵𝘁𝘀 𝗮𝗿𝗲 𝗻𝗼𝘄 𝗼𝗽𝗲𝗻-𝘀𝗼𝘂𝗿𝗰𝗲𝗱 — a RARE native unified model on an MoE backbone (Only 3B active! Lightning Fast⚡) 📄 Tech Report: arxiv.org/abs/2605.12500 🤗 Daily Papers (Vote & Discuss): huggingface.co/papers/2605.1… 🤗 Models: huggingface.co/collections/s… 💻 Code: github.com/OpenSenseNova/Sen… 🎮 Demo: unify.light-ai.top 👾 Discord: discord.com/invite/BuTXPHmQu…在 X 查看被引用的帖子
来源：@pumpkherm · x.com