[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mamoda-2-5-dit-moe-multimodal-95x":3,"topics-all":36,"news-related-f76fc8eb-1cd0-4577-aca8-db98018194ae":55},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"f76fc8eb-1cd0-4577-aca8-db98018194ae","Mamoda2.5：DiT-MoE 架构统一多模态生成，推理速度提升 95 倍","多模态大模型正在从「割裂理解与生成」走向「统一架构」。arXiv 近日公开的 Mamoda2.5 论文展示了这一方向的最新进展：其提出一种融合自回归（AR）理解模块与 Diffusion Transformer（DiT）生成模块的统一框架，首次在单一架构内同时完成多模态理解与内容生成。\n\n**技术核心：DiT-MoE 架构**\n\nMamoda2.5 的最大看点是将 Mixture-of-Experts（MoE）引入 Diffusion Transformer。其采用 128 路专家 Top-8 路由机制，总参数规模 250 亿，但每次推理仅激活 30 亿参数—— sparsity 设计使算力成本大幅压缩，同时保留了足够的模型容量。值得注意的是，团队还采用了从 Dense 模型到 MoE 的 upcycling 初始化策略，加速收敛。\n\n**视频编辑：超越 Kling O1**\n\n在 VBench 2.0 基准测试中，Mamoda2.5 达到顶级生成性能，尤其在视频编辑任务上刷新纪录，在 OpenVE-Bench 中超越快手 Kling O1。传统视频编辑模型通常需要 30 步迭代，而 Mamoda2.5 通过联合少步蒸馏与强化学习（joint few-step distillation + RL），将推理步骤压缩至 4 步，推理速度提升 95.9 倍，已在实际广告场景中落地，内容审核与创意修复成功率高达 98%。\n\n**评论**\n\nMamoda2.5 的意义在于三条技术主线的交汇：MoE 的稀疏化压缩、Diffusion Transformer 的统一生成范式、以及少步蒸馏的推理加速。这与业界追求的「更强能力 + 更低延迟」方向高度一致，也预示着 2026 年多模态模型竞争的核心将不在于参数规模，而在于稀疏路由效率和蒸馏压缩比。中小团队借助这类高效架构，有望在特定垂类场景逼近闭源大模型的体验。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.02641","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":21,"name":22,"slug":22,"description":13,"color":13},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"d08892be-aba4-4d62-bde3-4f7802cd82d5","en","Mamoda 2.5: DiT-MoE unifies generation, 95x faster inference","Multimodal large models are moving from \"split understanding and generation\" toward \"unified architecture.\" A recent Mamoda2.5 paper on arXiv showcases the latest progress in this direction: a unified framework fusing an autoregressive (AR) understanding module with a Diffusion Transformer (DiT) generation module — for the first time, multimodal understanding and content generation within a single architecture.\n\n**Technical core: DiT-MoE architecture**\n\nMamoda2.5's biggest highlight is the introduction of Mixture-of-Experts (MoE) into Diffusion Transformer. It uses a 128-expert Top-8 routing mechanism, with 25B total parameters but only 3B activated per inference — the sparsity design significantly compresses compute cost while preserving enough model capacity. Notably, the team also adopts an upcycling initialization strategy from Dense models to MoE, accelerating convergence.\n\n**Video editing: surpassing Kling O1**\n\nOn the VBench 2.0 benchmark, Mamoda2.5 achieves top-tier generation performance, particularly setting new records on video editing tasks, surpassing Kuaishou's Kling O1 on OpenVE-Bench. Traditional video editing models typically need 30 iterative steps, while Mamoda2.5, through joint few-step distillation + RL, compresses the inference steps to 4, boosting inference speed 95.9×. It has already been deployed in real advertising scenarios, with content review and creative restoration success rates as high as 98%.\n\n**Commentary**\n\nMamoda2.5's significance lies in the intersection of three technical threads: MoE's sparse-compression, Diffusion Transformer's unified generation paradigm, and few-step distillation's inference acceleration. This is highly consistent with the industry's pursuit of \"stronger capability + lower latency,\" and signals that the core of multimodal model competition in 2026 won't be parameter scale, but sparse-routing efficiency and distillation compression ratio. With the help of such efficient architectures, small and mid teams are poised to approach closed-source large-model experience in specific vertical scenarios.","mamoda-2-5-dit-moe-multimodal-95x","2026-05-18T13:10:00Z","2026-05-18T13:05:14.775823Z","2026-08-19T02:08:40.142862Z",true,"agent",256,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"dc534f1a-dbe2-4f79-9557-bbb43a82f24f","视频模型总在绕过 MLLM,三段训练把它拉回来","fapl-mllm-latents-cinematic-video-generation","2026-10-23T09:20:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"4349edac-997e-4980-95a0-58054a28a9dd","视频生成投机去噪:1.7 步顶 4 步,提速 1.5 倍","speculative-forcing-video-denoising","2026-10-16T19:11:48+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"0b0af156-7b5c-47d8-bfc2-de8cc0b22ce2","混元开源Prism:稀疏注意力让2K音视频训练提速2.5倍","prism-2k-video-audio-sparse-attention","2026-10-13T23:13:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"24d56562-4886-4169-8aa9-773f5fa34e2f","GRACE 把视频潜空间压 8 倍:提速 11 倍,跑分反超","grace-latent-compression-video-diffusion","2026-10-08T15:10:59+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"03d88e2a-526c-4c39-b9b6-e90ac0ff65e0","PixelUMM 拆掉 VAE:NVIDIA 用 8B 单 Transformer 在像素空间统一图文视频","pixelumm-encoder-free-unified-multimodal-8b","2026-09-29T21:59:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"af21d26d-d9cd-4526-ac4a-66366a45848c","AV-GRPO:8张A800给22B音视频模型做RL后训练","av-grpo-audio-video-diffusion-rl","2026-09-27T17:08:13+00:00"]