[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"topic-meta-h3-series":3,"topic-articles-h3-series":13},{"slug":4,"tag_slug":4,"title_zh":5,"title_en":6,"intro_zh":7,"intro_en":8,"id":9,"is_active":10,"created_at":11,"modified_at":12},"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",true,"2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"total":14,"page":15,"size":16,"items":17},7,1,100,[18,50,76,98,119,139,162],{"id":19,"title":20,"summary":21,"tags":22,"translations":39,"news_slug":46,"published_at":47,"created_at":48,"image_url":26,"view_count":49},"02c8b500-ec11-44a6-8c58-6e880563dad8","FastH3 开源:4 步蒸馏版 MiniMax H3,B200 单卡最高提速 14 倍","UCSD FastVideo 团队联合 NVIDIA FastGen、Nuva Lab 开源 FastH3 Preview v1:用 DMD2 蒸馏把 MiniMax H3 的 49 次 DiT 调用压到 4 次,叠加 90% 稀疏注意力,B200 单卡最高提速 14 倍,8 卡 15 秒视频不足 13 秒生成。",[23,27,30,33,36],{"id":24,"name":25,"slug":25,"description":26,"color":26},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":28,"name":29,"slug":29,"description":26,"color":26},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":31,"name":4,"slug":4,"description":32,"color":26},"f3be854a-50f0-411c-893d-16d0df6def02","MiniMax H3 专题：持续追踪 H3 的发布、开源、蒸馏与部署全链路",{"id":34,"name":35,"slug":35,"description":26,"color":26},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":37,"name":38,"slug":38,"description":26,"color":26},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[40],{"id":41,"lang":42,"title":43,"summary":44,"content":45},"8aabe7c4-d409-4361-a7c0-5c3abc0a998a","en","FastH3 Open-Sources a 4-Step MiniMax H3: Up to 14x on One B200","FastVideo open-sources FastH3 Preview v1: DMD2 distillation cuts MiniMax H3's 49 DiT calls to 4, plus 90% sparse attention — up to 14x on one B200.","The open-source video generation stack just gained a second layer. Weeks after MiniMax released the H3-Base weights, the FastVideo team at UCSD, working with NVIDIA's FastGen team and Nuva Lab, shipped FastH3 Preview v1 — a family of open-weight checkpoints post-trained on MiniMax H3 that push text-to-audio-video (T2VA) generation into a new latency class.\n\n## From 49 transformer calls down to 4\n\nThe cost problem is stark. Base H3's audio-video diffusion transformer has 33B parameters, and generating one video invokes it 49 times. FastH3 attacks both the number of calls and the work inside each call.\n\nThe first cut is DMD2 (Distribution Matching Distillation). A frozen Base H3 teacher and a learned critic train the student; the difference between their score estimates supplies the training signal, and backward simulation exposes prompt-only students to the few-step states they will see at inference. The result: five scheduler points, exactly four DiT forwards at generation time.\n\nThe second cut is VSA (Video Sparse Attention). The trainable sparse-attention student keeps about 10% of eligible video-to-video tiles (90% sparsity, 64-token blocks) while text and audio stay dense; the teacher and critic also run dense attention, giving the sparse student a full-attention target.\n\n## The numbers: up to 14.38x on a single B200\n\nAt 1344×768, 24 FPS with audio, Base H3 with dense FA4 takes 678.7 seconds for a 15-second clip on one B200; the recommended VSA\u002FData-Free checkpoint takes 47.2 seconds — a 14.38x speedup. A 5-second clip: 132.5s vs 16.2s, 8.16x. On 4× B200 a 5-second clip lands at 6.1 seconds end-to-end; on 8× B200 a 15-second clip lands at 12.88 seconds — faster than the clip itself plays. Timings are the median of three requests after warmup, with model load and compilation excluded, and end-to-end includes encoding, denoising, decoding, audio, muxing, and file output.\n\nTraining consumed 1k+ B200 hours. The recommended checkpoint was trained to step 1300 and ships both as full weights and as a pre-extracted rank-64 LoRA.\n\n## Data-free training, stated limits\n\nThe recommended checkpoint is data-free: trained from prompts only, without target videos. That sidesteps video-data sourcing and licensing headaches entirely. The team also publishes synthetic-data ablations and a dense-attention reference — four checkpoints in the Preview v1 family, with training code promised soon.\n\nThe limits are stated plainly: T2VA only (FL2VA and Ref2VA are not distilled yet), difficult motion, fine detail, and some audio may fall below Base H3 quality, and everything inherits the MiniMax H3 Community License.\n\n## Why this matters beyond \"another speedup\"\n\nIt validates the two-layer innovation model that open weights enable: MiniMax shipped the base, and the community now inspects it, post-trains it, swaps kernels, and runs it on its own hardware. fal's H3 Max, released the same week, took the hosted-API route — the two paths now have a direct comparison.\n\nDistillation plus sparse attention is becoming the standard route to production video generation. 49 calls down to 4, plus 90% attention sparsity, means the same hardware serves far more requests — or the same task fits into much smaller machines. Next on the roadmap: an 8-step quality variant, FP8\u002FNVFP4 quantization, local optimizations for RTX, DGX Spark, and Apple MLX, and a new Parallel Decoding Distillation (PDD) collaboration with NVIDIA.\n\nIf you are evaluating open video stacks, run the recommended VSA\u002FData-Free checkpoint: the weights are hardware-independent, the 4× B200 setup is a controlled benchmark rather than a requirement, and the GPU count just needs to divide H3's 56 attention heads.\n\nWhen generation time drops below the video's own duration, interactive and real-time workflows stop being demos and start being schedulable work. That is the real weight of this release.\n\nSource: https:\u002F\u002Fhaoailab.com\u002Fblogs\u002Ffasth3-preview\u002F","fasth3-4-step-distilled-minimax-h3","2026-08-30T21:30:00Z","2026-08-30T21:12:36.694350Z",396,{"id":51,"title":52,"summary":53,"tags":54,"translations":66,"news_slug":72,"published_at":73,"created_at":74,"image_url":26,"view_count":75},"1bf979b5-2b52-4a19-8e2d-e0a549bb24ba","fal 后训练版 MiniMax H3:5 秒视频约 3 秒生成,吞吐 35 倍","fal 在开源 MiniMax H3 权重上后训练出 H3 Max:5 秒视频约 3 秒生成,吞吐约为官方端点 35 倍,人工偏好评测三项居首,Design Arena 独立评测也称其速度超 50 倍;但分辨率止步 768p,仅支持文生视频与图生视频。",[55,56,59,62,65],{"id":31,"name":4,"slug":4,"description":32,"color":26},{"id":57,"name":58,"slug":58,"description":26,"color":26},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":60,"name":61,"slug":61,"description":26,"color":26},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":63,"name":64,"slug":64,"description":26,"color":26},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":37,"name":38,"slug":38,"description":26,"color":26},[67],{"id":68,"lang":42,"title":69,"summary":70,"content":71},"e6cdd454-5306-462a-9769-534f04040177","fal's H3 Max: post-trained MiniMax H3 makes 5s clips in 3s","fal's H3 Max: post-trained MiniMax H3 makes 5s clips in ~3s, ~35x official throughput, #1 in preference and Design Arena evals; capped at 768p.","Generative video has long run on an assumed tradeoff: a higher-quality model must be slower at inference. On August 27, inference platform fal challenged that assumption head-on with H3 Max, a post-trained version of MiniMax's open-weight H3 model. It generates a 5-second video in roughly 3 seconds — about 35x the throughput of the official MiniMax H3 endpoint — while ranking #1 in overall quality, prompt understanding, and aesthetics in fal's human preference evaluations ([fal blog](https:\u002F\u002Fblog.fal.ai\u002Fintroducing-h3-max-by-fal\u002F)).\n\n## Not a distillation copy, but model-and-inference co-design\n\nThe fal team started from the open-weights MiniMax H3 and introduced substantial new data during post-training, focused on prompt adherence and visual quality. According to fal's official X posts, a large portion of that post-training compute went to verifiable RL tasks. Unlike traditional step-distillation, which aims to match the base model, fal says it aimed for a better model at much faster speed — with the base model's core capabilities (unified multimodal context, natively synchronized audio and video) intact.\n\nThe inference side matters just as much. H3 Max's architecture was designed around fal's in-house inference engine — a team that has spent four years optimizing diffusion and generative-media workloads. Training and serving ran entirely on NVIDIA GB200 NVL72 systems, which fal says deliver up to 2x the per-chip performance of the previous-generation accelerators. Crucially, fal set a hard rule: common speedups such as reduced precision or fewer sampling steps only survived if the model still held its position in internal quality evaluations. Optimization and quality were treated as one problem, not two.\n\n## The leaderboard: two independent third parties agree\n\nfal benchmarked H3 Max against twelve leading video models in head-to-head human preference studies — including the official MiniMax H3 endpoint, Gemini Omni Flash, Wan 3.0, Seedance 2.5, Kling 3, and Veo 3.1 — aggregated with Bayesian Elo ratings and 95% confidence intervals. H3 Max ranked #1 across all three dimensions and won the majority of matchups against every model tested.\n\nThe claim doesn't rest solely on fal's own evaluations. In independent benchmarks from Artificial Analysis and Design Arena, H3 Max also ranks #1 among video models. Design Arena's wording: \"MiniMax H3 Max by fal delivers the quality of MiniMax H3 at more than 50x the speed,\" establishing what it calls a new speed–preference Pareto frontier. fal's own Pareto chart shows H3 Max generating in about 2.4 seconds while MiniMax H3 and Wan 3.0 take over a minute. The MiniMax H3 team also endorsed the work in fal's post, saying the two companies had \"worked closely ... since day one.\"\n\n## Read the fine print: a 768p ceiling and a misleading name\n\nSpeed has a price. The third-party page Morphic notes that H3 Max is tuned for speed rather than maximum resolution — it stops at 768p where the original H3 reaches 2K — and covers only text-to-video and image-to-video, dropping the reference generation and editing endpoints that mix images, clips, and audio. The name itself deserves a caveat: it is easy to read \"H3 Max\" as an official MiniMax tier, but MiniMax has never announced a model by that name. It is fal's tuned edition, hosted exclusively on fal.\n\nOne more overlooked comparison: fal's own product FAQ states that the standard H3 on fal's stack runs about 15x faster than MiniMax's own inference. In other words, a large share of the 35x gain comes from post-training and model-system co-design — not merely from switching hosting providers.\n\n## So what\n\nH3 Max matters beyond a single product launch. It validates the \"second-order value\" of open weights: by open-sourcing H3, MiniMax gained more than goodwill — infrastructure vendors like fal built a speed-quality combination on top that the official endpoint doesn't offer. The \"quality must be slow\" assumption now has counter-evidence; the right lens for model selection is no longer an isolated benchmark score but the frontier across quality, latency, and cost. The 50%-off first week signals that fal is chasing high-volume production workloads, not demos. As video generation moves toward interactive and high-volume production, \"treat model research and inference optimization as the same problem\" may well be the dividing line for the next generation of generative-media companies.","fal-h3-max-post-trained-video","2026-08-29T23:02:12Z","2026-08-29T23:05:26.363369Z",358,{"id":77,"title":78,"summary":79,"tags":80,"translations":88,"news_slug":94,"published_at":95,"created_at":96,"image_url":26,"view_count":97},"aad00b18-d354-48b5-ad21-62b53150b8c6","MiniMax H3 开源实测:你下载的权重,和 API 里跑的不是同一个模型","MiniMax H3 权重 8 月 3 日登陆 Hugging Face 后,社区实测发现:开源版只包含 API 三步流水线中的 base generation 模块,提示词增强与 2K 超分仍是托管服务,稀疏注意力实现也未随首发放出。本地跑通原生 768p 之外,MLX 移植、Turbo-LoRA 约 5 倍采样加速、提示词重写 LoRA 等社区补完速度惊人。本文拆解本地与 API 的真实差距、显存门槛与许可证边界。",[81,84,85,86,87],{"id":82,"name":83,"slug":83,"description":26,"color":26},"fca9258a-9430-455a-b95d-b9fae5e373a8","ai-inference",{"id":31,"name":4,"slug":4,"description":32,"color":26},{"id":63,"name":64,"slug":64,"description":26,"color":26},{"id":34,"name":35,"slug":35,"description":26,"color":26},{"id":37,"name":38,"slug":38,"description":26,"color":26},[89],{"id":90,"lang":42,"title":91,"summary":92,"content":93},"7f7b85ab-41b2-41ca-b666-461fba027832","MiniMax H3 tested: open weights differ from the API model","After MiniMax H3's weights landed on Hugging Face on August 3, community testing revealed that the open-source release only ships the base generation block of the API's three-step pipeline: prompt enhancement and 2K upscaling remain hosted services, and the sparse-attention implementation is withheld for a future update. Beyond native 768p local runs, the community response has been remarkably fast, including an MLX port, a Turbo-LoRA with roughly 5x sampling speedup, a prompt-rewriter LoRA, and a live LoRA fine-tuning trainer. This piece breaks down the real gap between local and API, the VRAM ladder, and the license boundaries.","MiniMax H3 Open-Weights Field Test: What You Download Is Not Quite What the API Runs\n\nOn August 3, MiniMax put the weights of H3 on Hugging Face. It is a rare heavyweight open release for video generation: a 33B-parameter diffusion transformer with Qwen3-VL-32B as the encoder, outputting audio-video clips up to 15 seconds long with native stereo sound. One week after the release, community testing delivered a calmer verdict — what you download to your machine and what the API runs are, strictly speaking, not the same system.\n\n## The open release ships only one block of the API stack\n\nOxen.ai's teardown puts it plainly: behind the H3 API is a three-step pipeline — first enhance the user's prompt, then run the base model, then run an upscaler to reach 2K resolution. The current open-source release contains only the base generation block. To reproduce API-level results locally, you have to assemble the other two steps yourself.\n\nThe official Hugging Face model card confirms this: H3-Context-IR, the system responsible for multimodal preprocessing and orchestration, is a hosted service and not in the weights package; native resolution locally defaults to 768p, while 2K is only available through the hosted API. More importantly, the model card states explicitly that H3 natively supports sparse-attention training and inference, but the initial open-source release provides full-attention inference only, with the sparse implementation to arrive in a future update — the compute savings on long multimodal sequences are, for now, out of reach for local users.\n\n## The VRAM ladder and the speed of community patching\n\nLocal requirements are more stratified than expected: Oxen.ai's guidance is at least 24 GB of VRAM; DiffSynth-Studio's NF4 quantization with offloading squeezes it onto as little as 8 GB; for higher-precision weights and longer generations, 50-60 GB is the sweet spot. The author's own setup is two 48 GB A6000s, with the DiT on one GPU and the text encoder plus VAEs on the other.\n\nThe real surprise is the community's speed. Only about a week in, nearly every layer of the stack is being rewritten: Simon Willison produced an MLX port for Apple Silicon; antirez wrote a C library and command-line tool on top of Apple's Metal; larryvrh's Turbo-LoRA cuts sampling from the usual ~20 steps down to 4, a roughly 5x sampling speedup; lightx2v's prompt-rewriter LoRA is effectively a local stand-in for the API's prompt-enhancement step; and Oxen.ai's LoRA fine-tuning trainer is already live.\n\n## Do not skip the license map\n\nAnother thing easily obscured by the word open: the license terms. H3 ships under the MiniMax Community License which, as reported by the South China Morning Post, permits free non-commercial use and commercial use for organizations with under 20 million US dollars in annual revenue, with attribution requirements. ExplainX's analysis goes further, noting that the license excludes markets including the United States, the European Union, the United Kingdom and South Korea — production use in those regions needs MiniMax's separate commercial licensing path. The first step in evaluating this model is not the benchmark scores; it is where your company sits on that map.\n\n## So what\n\nOxen.ai's verdict: it finally feels like having a Seedance 2.0-level model at home. Yet the distance between downloadable and reproducible is an entire pipeline wide — that is equally a fact. The signal to watch is clear: as the sparse-attention implementation, the upscaler, and local stand-ins for prompt enhancement get filled in layer by layer, H3 turns from an open model you can run into an open system you can use. On that day, the closed-source moat around video generation is genuinely dismantled.\n\nSource: https:\u002F\u002Fwww.oxen.ai\u002Fblog\u002Fminimax-h3","minimax-h3-local-vs-api-gap","2026-08-15T17:07:24Z","2026-08-15T17:09:11.145823Z",187,{"id":99,"title":100,"summary":101,"tags":102,"translations":108,"news_slug":114,"published_at":115,"created_at":116,"image_url":117,"view_count":118},"6e3002da-c1fd-4a6d-b903-4f65b976dd04","MiniMax H3 首个商用落点：美图 RoboNeo 接入背后,通用多模态模型的\"可编辑性\"才刚开始被检验","8 月 3 日,美图旗下 RoboNeo 宣布接入 MiniMax 刚开源的多模态生成模型 MiniMax H3,主打多模态理解与精细化视频局部编辑。同日,摩尔线程基于 MTT S5000 完成 H3 的 Day-0 适配。从开源到商用、从训练到推理,H3 正在被打磨成一款真正可用的多模态生产力工具。",[103,104,105,106,107],{"id":31,"name":4,"slug":4,"description":32,"color":26},{"id":60,"name":61,"slug":61,"description":26,"color":26},{"id":63,"name":64,"slug":64,"description":26,"color":26},{"id":34,"name":35,"slug":35,"description":26,"color":26},{"id":37,"name":38,"slug":38,"description":26,"color":26},[109],{"id":110,"lang":42,"title":111,"summary":112,"content":113},"f8ff89df-abfd-4456-9bc0-91edab3399e6","MiniMax H3's first landing: Meitu RoboNeo tests editability","On August 3, Meitu's RoboNeo announced its integration with MiniMax's newly open-sourced multimodal generation model MiniMax H3, focusing on multimodal understanding and fine-grained video editing. The same day, Moore Threads completed Day-0 adaptation of H3 on its MTT S5000 card. From open source to commercial use, from training to inference, H3 is being polished into a truly usable multimodal productivity tool.","## RoboNeo Catches the First Commercial Pitch from MiniMax H3\n\nOn August 3, Meitu's RoboNeo announced a formal integration with MiniMax H3, strengthening multimodal understanding and fine-grained local video editing capabilities (source: 36Kr report on Meitu's RoboNeo integrating MiniMax H3). At a time when multimodal generation models are still racing on parameter counts and resolution, a real video editing product has chosen to use H3 as its backbone — a signal worth digging into.\n\nAccording to 36Kr's disclosure, after integrating H3, RoboNeo can do the following:\n\n- **Unified multimodal understanding**: text, image, video, and audio processed in a single frame; the model treats them as one unified generation input internally.\n- **Fine-grained local video editing**: character swap, object add\u002Fremove, background change, effect tweak, voice tone transfer, line rewriting.\n\nBoth capabilities directly map to MiniMax's H3 core selling points listed in its official blog — \"a general-purpose omni-modal generation model,\" one model handling both understanding and generation across text, images, video, and audio (source: MiniMax Blog · MiniMax H3, 2026-08-03).\n\nIn other words, H3 is not just another \"generate 8 seconds of pretty clips\" video model — it wants to become generative-plus-editing infrastructure that can run inside real workflows.\n\n---\n\n## Why H3's Design Philosophy Determines This Kind of Usage\n\nMiniMax spelled out H3's design principle pretty bluntly on the official blog: the previous two generations (Hailuo 01, Hailuo 02) split tasks apart — separate expert models for image, video, audio. H3 reverses that and **converges** tasks into a single generative framework.\n\nThree specific anchors:\n\n1. **H3-Omni Transformer** replaces Hailuo 02's specialized architecture. The authors explicitly say that because H3 does task generalization, those prior architectural tricks became \"unnecessary complexity.\"\n2. **Early training-stage fusion** of modalities and tasks, with a dedicated captioning pipeline that compresses roughly 100K tokens of raw material down to an average of 4K tokens of \"contextual omni representation.\"\n3. **2K output via in-context regeneration** — no external super-resolution module. The H3 base model upscales its own low-resolution output within the multimodal context, recovering small text and fine detail noticeably better than traditional upscaling.\n\nThis architecture, viewed against today's multimodal competition, has a **most-direct byproduct: controllable editing.** RoboNeo's character swap, line rewrite, background change — all are \"on existing video, take a natural-language instruction, modify one piece\" — exactly the kind of task H3's \"unified tasks + language as the universal bridge\" line was designed for.\n\n---\n\n## Same Day: Moore Threads Completes Day-0 Adaptation\n\nAnother thing happening on the same day: Moore Threads, based on its AI training-inference integrated card MTT S5000 and its MUSA software stack, completed Day-0 adaptation and runtime for H3 (source: 36Kr report, 2026-08-03).\n\nThis is easy to overlook but not insignificant:\n\n- **Domestic inference stacks are starting to treat \"multimodal large models\" as a category of infrastructure to adapt**, not something each application layer reinvents from scratch.\n- MTT S5000 + MUSA being runnable within a day means H3 was actively designed for hardware compatibility — MiniMax stated on the blog, \"hardware compatibility has been a key consideration since the earliest stages of H3's design.\"\n\nStacked together, these events show H3 building a cadence of **\"model open-source + inference stack Day-0 + application-layer Day-1 integration.\"** Fancy model launches are common, but having \"zero-day response\" simultaneously on open source, hardware, and application layers — in the Chinese AI community today that is rare.\n\n---\n\n## Commercial Feedback Becomes H3's Litmus Test\n\nFrom training paradigm to product deployment sits a long ditch. Once RoboNeo and similar P-end products start using H3, certain issues will inevitably surface, and those issues will redefine the priorities of H3's next versions:\n\n- **Controllability vs. stylization**: commercial scenarios demand extreme stability around \"change one piece, don't touch the rest.\" Whether H3's \"natural-language instructions + in-context regeneration\" line holds up under the consistency-sensitive demands of e-commerce assets and ad voice-overs is the key test.\n- **2K default vs. inference cost**: MiniMax says H3's per-second price at 2K is less than a third of mainstream models, and at 768p is less than half of the mainstream 720p price — that is relative, not absolute. Once commercial traffic ramps, inference TCO will be the second number repeatedly pulled up in spreadsheets.\n- **Multimodal context length**: with 15-second video the context is roughly 100K tokens compressed to 4K — how that compression ratio plays out will be easy to quantify in RoboNeo's historical-asset style workflow.\n\nSo this integration is not, for MiniMax, \"another customer used us\" prestige — it is putting H3's task-generalization claim under real industrial stress test.\n\n---\n\n## So What: The Second Half of Multimodal Generation Hinges on Who Gets Production-Used First\n\nIn 2026's multimodal generation race, everyone is pushing on native resolution, context length, or hardware \u002F model compatibility. H3 has taken a more visible but harder path — **\"one model for all tasks + open source + 2K default** — with the commercial bet squarely on \"editability.\"\n\nRoboNeo, ComfyUI Day-0 support, Moore Threads Day-0 adaptation — that streak puts H3 inside real workflows of \"open-source model + inference stack + editing product\" within a single week. What to watch next is not how many more demos MiniMax releases, but whether real products like RoboNeo can **actually replace one specific segment of the traditional video editing pipeline** after one or two months of use.\n\nIf yes, H3 is the one that genuinely moves multimodal generation from \"demo\" to \"productivity.\" If not, Veo, Kling, Sora, and the queue behind them will quickly fill the same gap.\n\n**References:**\n- Meitu's RoboNeo integrates MiniMax H3, 36Kr, 2026-08-03 (https:\u002F\u002Fwww.36kr.com\u002Fnewsflashes\u002F3923674190999173)\n- Moore Threads completes MiniMax H3 multimodal model adaptation, 36Kr, 2026-08-03 (https:\u002F\u002Fwww.36kr.com\u002Fnewsflashes\u002F3923276079263113)\n- MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities, MiniMax Blog, 2026-08-03 (https:\u002F\u002Fwww.minimax.io\u002Fblog\u002Fminimax-h3)","roboneo-minimax-h3-multimodal-editing","2026-08-03T18:02:02Z","2026-08-03T18:02:17.425224Z","https:\u002F\u002Ffile.cdn.minimax.io\u002Fpublic\u002F3df321d9-42bd-4be0-ac58-71f22377a11f.png",207,{"id":120,"title":121,"summary":122,"tags":123,"translations":129,"news_slug":135,"published_at":136,"created_at":137,"image_url":26,"view_count":138},"a970be54-0603-4446-b92e-67c887c0788a","ComfyUI Day-0 跑通 MiniMax H3:把 2K 多模态视频生成压到 RTX 3060","MiniMax 开放权重发布 H3(Hailuo 3.0),ComfyUI 在发布当天通过 66% 显存压缩把 2K 多模态视频生成跑进消费级 GPU;Artificial Analysis 榜单显示 H3 在视频编辑方向排名第一,文生视频与图生视频分列第二、第三。",[124,125,126,127,128],{"id":82,"name":83,"slug":83,"description":26,"color":26},{"id":31,"name":4,"slug":4,"description":32,"color":26},{"id":63,"name":64,"slug":64,"description":26,"color":26},{"id":34,"name":35,"slug":35,"description":26,"color":26},{"id":37,"name":38,"slug":38,"description":26,"color":26},[130],{"id":131,"lang":42,"title":132,"summary":133,"content":134},"1c3a10f9-19c0-4a38-8989-e3f39d6c289c","ComfyUI runs MiniMax H3 on day one: 2K video generation on RTX 3060","On 3 August 2026 MiniMax released the open-weights MiniMax H3 (Hailuo 3.0) omni-modal video model, and ComfyUI shipped Day-0 support that compresses the 2K inference memory footprint by 66% so the model runs locally on a 12 GB RTX 3060. The Artificial Analysis leaderboard puts H3 at #1 in video editing, #2 in text-to-video, and #3 in image-to-video.","# ComfyUI Day-0 support for MiniMax H3: a 2K omni-modal video model that fits on an RTX 3060\n\n> Hook: On the same day a 2K-resolution, stereo-audio, open-weights omni-modal video model ships, ComfyUI lands a working local-inference path on a 12 GB RTX 3060. That is the first real desktop route for \"production-grade\" AI video.\n\n## Background: from Hailuo to H3, the open-weights moment\n\nOn 3 August 2026, Shanghai-based MiniMax (MiniMax Group Inc.) officially released the third-generation omni-modal generation model **MiniMax H3** (a.k.a. Hailuo 3.0) as an open-weights release. MiniMax's blog describes H3 as \"a general-purpose omni-modal generation model that can jointly understand multimodal contexts spanning text, images, video, and audio\" and generate video with \"native stereo audio at up to 2K resolution and 15 seconds in length\" ([MiniMax blog](https:\u002F\u002Fwww.minimax.io\u002Fblog\u002Fminimax-h3)). H3 is the third video model in the Hailuo lineage (after Hailuo 01 and Hailuo 02) and the first that MiniMax has released with open weights. The model card is live on Hugging Face ([Hugging Face: MiniMaxAI\u002FMiniMax-H3](https:\u002F\u002Fhuggingface.co\u002FMiniMaxAI\u002FMiniMax-H3)).\n\nThe third-party evaluation house Artificial Analysis posted a leaderboard on X showing H3 ranking #1 in Video Editing, #2 in Text-to-Video, and #3 in Image-to-Video ([Artificial Analysis on X](https:\u002F\u002Fx.com\u002FArtificialAnlys\u002Fstatus\u002F2083042088338538594)). The Decoder wrote on 3 August 2026 that MiniMax H3 was \"the first open model to top an AI video ranking\" ([The Decoder](https:\u002F\u002Fthe-decoder.com\u002Fchinas-minimax-h3-is-the-first-open-model-to-top-an-ai-video-ranking\u002F)).\n\nThe same day, Comfy Org published \"MiniMax H3 Day-0 Support in ComfyUI\" on its Substack newsletter, detailing the engineering work that brought H3 into ComfyUI 0.30.0 and the optimizations that made local inference feasible on consumer hardware ([Comfy Org blog](https:\u002F\u002Fblog.comfy.org\u002Fp\u002Fminimax-h3-day-0-support-in-comfyui)). Two numbers from the post matter most: by pruning roughly 40% of the model parameters (the \"modulation weights\") into a functionally equivalent lookup table, layering an int8 convrot quantization on top, and shipping custom inference kernels, the Comfy team pushed the total memory footprint of the smallest H3 variants from 123.6 GB in full precision down to 42.5 GB — a 66% reduction. Combined with dynamic VRAM offloading, the result is a 2K-class omni-modal video model that runs locally on an RTX 3060.\n\n## What ComfyUI actually did\n\nThe ComfyUI integration can be broken into three engineering moves, each of which directly affects \"can I run this at home\":\n\n- **Multimodal context understanding** is the headline capability. H3 takes images, audio, and video in one prompt and resolves them against a description of how the inputs relate, collapsing what would otherwise be five separate tasks (script + image references + audio + reference video + edit pass) into a single model call. The implication: H3 is not a T2V model that happens to accept an image, it is a different architectural choice about where the cross-modal work happens.\n- **Native stereo audio** is a model property, not a post-process. Every audio output is generated in stereo from the same forward pass that produces the video — there is no separate TTS or audio model bolted on afterwards. The Decoder notes that a single H3 prompt can include up to 9 reference images, 3 reference video clips, and 3 reference audio clips ([Hugging Face model card](https:\u002F\u002Fhuggingface.co\u002FMiniMaxAI\u002FMiniMax-H3)).\n- **Editing and motion transfer** support a reference-video-supplies-motion, style-comes-from-elsewhere workflow, and combined with in-place editing that means a single shot can be iterated without regenerating the whole clip. For a graph-style pipeline (ComfyUI's natural habitat), this is the capability that changes how workflows are built.\n\nComfyUI ships three official workflow templates: text-to-video (T2V), image-to-video (I2V), and reference-to-video (R2V) ([Comfy-Org workflow_templates](https:\u002F\u002Fgithub.com\u002FComfy-Org\u002Fworkflow_templates\u002Fblob\u002Fmain\u002Ftemplates\u002Fvideo_minimax_h3_i2v.json)). Weights are hosted at `Comfy-Org\u002FMiniMax-H3`; running locally requires ComfyUI 0.30.0 or later plus the model files in the standard ComfyUI directories.\n\nA practical limit to be aware of: ComfyUI's local H3 workflow caps at 768p. The Decoder reports the same ceiling. The 2K path depends on an **H3-Context-IR** module that turns prompts and reference material into a structured intermediate representation, and that module is not included in the open-weights release. The community is expected to assemble the context side themselves using MiniMax's published prompting guides.\n\n## Opinion: H3's engineering impact outranks \"yet another video model\"\n\nH3 lands at an interesting moment in the 2026 video-generation landscape, and three things stand out:\n\n**1. \"Native omni-modal\" is moving from concept to working pipeline.** Most video models in the past 12 months sold themselves on visual quality and motion stability, with audio and reference understanding handled by separate models. H3 takes four input modalities in one forward pass and emits synchronized stereo video. Whether the underlying mechanism is \"truly native\" or a tightly-coupled multi-encoder stack is something the open-weights release will let the community verify. What is clear from the ComfyUI Day-0 integration is that the API surface is self-consistent enough to build a real workflow on.\n\n**2. Open-weights 2K video has its first real \"desktop\" path.** ComfyUI's 66% memory reduction plus dynamic VRAM offloading is what moves \"consumer GPU running a 2K multi-modal video model\" from marketing slide to executable. The local ceiling is 768p, so production 2K output still requires either the closed 2K module from MiniMax or an upscale pass — but the bottleneck is no longer \"you need eight A100s to try the model.\" The same day, ByteDance released the closed **Seedance 2.5** with 30-second clips and built-in audio, which is a different point on the design space (\"long + closed\" vs. \"open + multi-modal + short\").\n\n**3. Chinese vendors are bundling \"multi-modal weights + infrastructure\" together.** The pattern repeats: 16 July 2026 saw Moonshot's Kimi K3 launch as a 2.8T-parameter open-weights model (\"the world's first open 3T-class model\" per Moonshot's own framing on [openlm.ai](https:\u002F\u002Fopenlm.ai\u002Fkimi-k3\u002F)); 19 July 2026 saw Alibaba open-weight the Qwen3.8-Max preview at 2.4T parameters ([Alibaba Cloud blog](https:\u002F\u002Fwww.alibabacloud.com\u002Fblog\u002Falibaba-cloud-unveils-agent-native-innovations-at-waic-2026_603377)); and on 3 August 2026 MiniMax H3 took the AI video leaderboard with a 33B open-weights release. The \"Chinese stack goes fully open\" story that started in LLMs is now visibly extending to multi-modal generation.\n\n## So what: what practitioners should actually do\n\nFor anyone shipping AI video products or building local creative tooling:\n\n- **Update ComfyUI to 0.30.0 and run all three H3 templates today.** T2V, I2V, and R2V cover roughly 80% of non-interactive use cases, and side-by-side comparison against Veo \u002F Kling \u002F Runway on the same prompt is the fastest way to locate where H3 still has visible gaps.\n- **For 2K commercial deliverables, wait for the missing pieces.** The 2K inference module and the H3-Context-IR are still closed. Until MiniMax releases them, the practical path is 768p for offline previews plus a separate upscale pass.\n- **Domain fine-tuning is feasible but commercially gated.** Fine-tuning H3 on a specific IP, character, or visual style is supported by the open weights, but the commercial-use license restricts it to companies under US$20M in annual revenue. Larger companies need to negotiate separately.\n\nThe second half of 2026 is no longer a one-dimensional race about \"who is longer, who is sharper.\" H3's three-part answer is \"omni-modal native + open weights + desktop-runnable.\" The next frontier is the intersection of \"open-weights 2K + complete context module\" and \"consumer-GPU long video.\"\n\n## References\n\n- MiniMax H3 blog: \u003Chttps:\u002F\u002Fwww.minimax.io\u002Fblog\u002Fminimax-h3>\n- Hugging Face model card: \u003Chttps:\u002F\u002Fhuggingface.co\u002FMiniMaxAI\u002FMiniMax-H3>\n- ComfyUI Day-0 engineering note: \u003Chttps:\u002F\u002Fblog.comfy.org\u002Fp\u002Fminimax-h3-day-0-support-in-comfyui>\n- Comfy-Org workflow templates: \u003Chttps:\u002F\u002Fgithub.com\u002FComfy-Org\u002Fworkflow_templates\u002Fblob\u002Fmain\u002Ftemplates\u002Fvideo_minimax_h3_i2v.json>\n- The Decoder coverage: \u003Chttps:\u002F\u002Fthe-decoder.com\u002Fchinas-minimax-h3-is-the-first-open-model-to-top-an-ai-video-ranking\u002F>\n- Artificial Analysis leaderboard: \u003Chttps:\u002F\u002Fx.com\u002FArtificialAnlys\u002Fstatus\u002F2083042088338538594>\n- lmmarketcap model timeline: \u003Chttps:\u002F\u002Flmmarketcap.com\u002Fllm-updates>\n- Kimi K3 framing: \u003Chttps:\u002F\u002Fopenlm.ai\u002Fkimi-k3\u002F>\n- Qwen3.8-Max preview: \u003Chttps:\u002F\u002Fwww.alibabacloud.com\u002Fblog\u002Falibaba-cloud-unveils-agent-native-innovations-at-waic-2026_603377>","comfyui-day0-minimax-h3-2k-video","2026-08-03T04:00:00Z","2026-08-03T14:43:58.146726Z",1853,{"id":140,"title":141,"summary":142,"tags":143,"translations":151,"news_slug":157,"published_at":158,"created_at":159,"image_url":160,"view_count":161},"83bf0960-2a51-4519-8326-6977527a68d9","MiniMax H3 三小时跑上 MTT S5000：Day-0 适配真正比拼的是软件栈","MiniMax H3 开源当天，摩尔线程称仅用 3 小时便在单机八卡 MTT S5000 节点完成稳定部署，串起 SGLang-MUSA、MATE、muDNN 与 MUSA Runtime。本文拆解这次 Day-0 适配的工程含义：国产 GPU 的竞争焦点，正从“能不能跑”转向框架兼容、算子优化和持续跟进模型架构的速度。",[144,145,148,149,150],{"id":31,"name":4,"slug":4,"description":32,"color":26},{"id":146,"name":147,"slug":147,"description":26,"color":26},"e0d31e94-ce47-4c8f-831c-d3d2926d42f3","hardware",{"id":57,"name":58,"slug":58,"description":26,"color":26},{"id":63,"name":64,"slug":64,"description":26,"color":26},{"id":37,"name":38,"slug":38,"description":26,"color":26},[152],{"id":153,"lang":42,"title":154,"summary":155,"content":156},"8ca4f22f-58aa-47ad-8e7b-fbb2c239fa34","MiniMax H3 runs on MTT S5000 in three hours: software stack wins","On the day MiniMax open-sourced H3, Moore Threads said it achieved stable deployment on a single eight-card MTT S5000 node in only three hours, linking SGLang-MUSA, MATE, muDNN, and MUSA Runtime. This article examines the engineering significance of that day-zero result and argues that competition among non-CUDA GPU platforms is shifting from basic compatibility toward framework integration, operator optimization, and the speed of supporting new model architectures.","# MiniMax H3 Runs on MTT S5000 in Three Hours: Day-Zero Support Is Really a Software-Stack Test\n\nHow quickly can a newly open-sourced model run on a computing stack outside the CUDA ecosystem? Moore Threads says its answer for MiniMax H3 was **three hours**.\n\nOn August 3, MiniMax released H3 as an open multimodal generation model. Moore Threads then reported that it had deployed the model and achieved stable operation on a single eight-card MTT S5000 node. The team connected the inference framework, operator libraries, compiler, and runtime into a complete execution path. The important part is not simply that another model can run on a domestic GPU. It is what this day-zero result reveals about the maturity of the surrounding software.\n\n## Why H3 Is Harder to Port Than a Conventional Language Model\n\nH3 accepts text, images, audio, and video as inputs. It can generate videos at up to 2K resolution, as long as 15 seconds, with native audio. According to Moore Threads, multimodal context increased the variance in H3's sequence lengths by roughly three times, while both the understanding and generation phases imposed substantially heavier compute loads.\n\nThat creates challenges beyond large matrix multiplications. The inference system must handle irregular input lengths, cross-modal data flows, and a video-generation pipeline. High theoretical throughput alone is insufficient: if the framework cannot schedule workloads efficiently, critical kernels are poorly implemented, or the runtime moves data too often, a model may technically start but still be unsuitable for a stable service.\n\n## Three Hours Was Not Magic; It Was the Payoff From Four Prepared Layers\n\nMoore Threads says its engineers decomposed the architecture, analyzed its core techniques, and identified representative operators before linking the following stack:\n\n- **SGLang-MUSA** serves as the model and inference-framework layer and was merged into the main SGLang project in April;\n- **SGLang-Diffusion and sgl-kernel** provide the multimodal generation subsystem and high-performance operator integration;\n- **MATE and muDNN** handle capability detection, kernel selection, backend scheduling, and optimized implementations of core operations such as Attention and GEMM;\n- **MTCC and MUSA Runtime** cover compilation and execution, completing the path from the upper-level framework to the GPU.\n\nIn other words, “three-hour adaptation” does not mean engineers built an entire backend after H3 appeared. It means previous investments in framework compatibility, operator interfaces, and runtime infrastructure could be reused immediately. Day-zero support is the compounding return on accumulated software assets.\n\n## What the Result Proves—and What It Does Not\n\nThe result shows that competition among Chinese GPU platforms is moving beyond the question of whether a model can run at all. The more relevant question is how quickly a new architecture can be integrated reliably. Model families now change on a weekly cadence. A hardware platform that requires months of custom work for each release will struggle to enter developers' real workflows, regardless of attractive peak specifications.\n\nHowever, the claim should be interpreted carefully. Moore Threads disclosed the adaptation time and operation on an eight-card node, but it did not publish end-to-end generation latency, throughput, memory use, power consumption, or a like-for-like comparison with other accelerators. This is therefore a meaningful **compatibility and engineering-efficiency milestone**, not a final verdict on performance. Reproducible benchmarks, concurrent-serving results, and long-duration stability data would be the stronger next evidence.\n\n## The Moat for Domestic Compute Will Look Increasingly Like a Software Company\n\nBecause H3 is open, enterprises can deploy it locally and customize it with proprietary data. A non-CUDA platform that supports the model on release day reduces the time cost of evaluation and migration. For GPU vendors, the contest will increasingly depend not only on chip area and theoretical compute, but on sustained participation in upstream ecosystems such as PyTorch and SGLang, minimal code changes for users, and the ability to reuse operator optimizations across successive models.\n\n**The chip defines the performance ceiling, but the software stack determines whether that chip reaches production in time. The real value of day-zero support is not merely that a model “runs”; it is that the wait for support can shrink from months to hours.**","minimax-h3-mtt-s5000-day-zero-stack","2026-08-02T19:43:00Z","2026-08-03T06:10:17.711815Z","https:\u002F\u002Fmmbiz.qpic.cn\u002Fmmbiz_jpg\u002FrnU4nRsEjiayMWULbicdEOrfcwGvS2icaVHYj3XgvOINriaKy1epb2UYV2ee1q628F4F0WS8IkVUGTvW73ovVBN5GH6l8uVoCzarZlCwl8CWOUo\u002F0?wx_fmt=jpeg",219,{"id":163,"title":164,"summary":165,"tags":166,"translations":174,"news_slug":180,"published_at":181,"created_at":182,"image_url":117,"view_count":183},"6f375936-79af-4622-a75e-d802ade563e0","MiniMax H3 不只是 2K 视频：它想把生成、参考和编辑收回一个模型","MiniMax 发布通用全模态视频模型 H3，把文本、图像、视频与音频统一为上下文，可生成最高 2K、最长 15 秒并带原生立体声的视频。它真正值得关注的不是规格，而是用 H3-VAE、Omni Transformer 与上下文再生成，把分裂的生成、参考和编辑任务收回一个模型。",[167,168,169,170,173],{"id":31,"name":4,"slug":4,"description":32,"color":26},{"id":60,"name":61,"slug":61,"description":26,"color":26},{"id":63,"name":64,"slug":64,"description":26,"color":26},{"id":171,"name":172,"slug":172,"description":26,"color":26},"4f214978-cac1-4f39-aa4b-f92a0d0934b7","transformer",{"id":37,"name":38,"slug":38,"description":26,"color":26},[175],{"id":176,"lang":42,"title":177,"summary":178,"content":179},"18d12800-c096-4cc6-a3c5-d46e5bb46217","MiniMax H3 unifies generation, reference, and editing","MiniMax has introduced H3, a general-purpose omni-modal video model that accepts text, images, video and audio in one context and generates up to 15-second 2K video with native stereo sound. The more important shift is architectural: H3 uses a redesigned VAE, an Omni Transformer and in-context regeneration to unify generation, reference and editing tasks that were previously split across specialized pipelines.","## Video generation is starting to move beyond one model per task\n\nFor the past two years, progress in video generation has been measured mainly by resolution, duration and motion quality. Yet production workflows remain fragmented: text-to-video, image-to-video, first-and-last-frame generation, subject reference, motion transfer, dubbing, sound effects and editing are often handled by separate models or pipelines. MiniMax H3 is designed to challenge that fragmentation. It accepts text, images, video and audio in a unified context and generates video at up to 2K resolution, 15 seconds in length, with native stereo sound.\n\nThat makes H3 more than another model with a 2K headline. Its central thesis is that **task boundaries have become a bottleneck**. A request such as ‘use the camera move from video one, make the person in image two sing, and match audio three’ normally has to be decomposed into multiple tools. H3 aims to understand the relationships among those inputs and carry out understanding, generation and editing inside one model.\n\n## Three technical choices matter more than the demos\n\nThe first is **Contextual Omni Representation**. MiniMax says its understanding pipeline describes not only the target video, but also the relationships among source assets and between the context and the target. A source item can require about 100,000 tokens of inference before being distilled into roughly 4,000 tokens on average. Language therefore acts as a general interface that represents tasks and cross-modal relationships, rather than merely serving as a prompt.\n\nThe second is a redesigned **H3-VAE**. Its higher compression ratio produces a fourfold gain in effective sequence length, lowering training and inference costs while enabling native 2K output. This is important because spatiotemporal token counts expand rapidly with resolution and duration. A video tokenizer that reconstructs well but is hard for the downstream model to learn still leaves the Transformer buried under long sequences. H3 treats the tokenizer as a core architectural component.\n\nThe third is the **H3-Omni Transformer** and its heterogeneous training system. Adding omni-modal context tripled the variance in sequence length, while understanding and generation created different compute profiles. MiniMax separated those workloads for hardware optimization and balanced heterogeneous compute across samples, reporting an end-to-end training throughput gain of nearly 30%. This is less spectacular than a demo reel, but it determines whether a video model can iterate economically.\n\n## In-context regeneration is a smarter approach to 2K\n\nInstead of attaching a conventional super-resolution network, H3 asks the base model to regenerate its own low-resolution result while seeing the original multimodal context again. Standard super-resolution can only infer details from existing pixels and often invents small text, logos or fine textures. Regeneration can revisit the original text, images, audio and reference video, giving the model a better chance to restore semantically correct details.\n\nMiniMax also claims that H3 costs less than one-third as much per second as mainstream models at 2K, and that its 768p output costs less than half as much as mainstream 720p generation. Those figures remain vendor claims. The full technical report has not yet been released, and the weights are promised for the coming days rather than already available. Until independent benchmarks, memory requirements and a concrete license appear, the open-model claim should be treated as incomplete.\n\n## The larger signal: video models are becoming production systems\n\nH3 matters less as a possible winner in a single quality comparison than as an attempt to reduce task switching. If unified pretraining can cover generation, reference, editing and synchronized audio-visual modeling, the competitive advantage of creative tools will shift away from chaining specialist models and toward organizing context, data and feedback around a general model.\n\nThe next things to watch are therefore not the showcase clips. They are whether the weights are actually released, whether third-party hardware can run the model efficiently, and whether character, text and brand consistency survive complex editing. The next stage of video generation may not be a longer clip; it may be one model taking responsibility for the entire creative pipeline.","minimax-h3-omnimodal-video-unified-generation-editing","2026-08-03T04:08:31Z","2026-08-03T04:14:53.550355Z",212]