[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-hyperflow-h3-flash-int8-consumer-gpu-video":3,"topics-all":39,"news-related-65fc5af8-80a4-4ba0-b9b9-6714d0968c25":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":25,"news_slug":32,"published_at":33,"created_at":34,"modified_at":35,"is_published":36,"publish_type":37,"image_url":14,"view_count":38},"65fc5af8-80a4-4ba0-b9b9-6714d0968c25","第三方压H3到13.7B:消费卡9.1GB权重跑768p","Video Rebirth 把 MiniMax H3 压到 13.7B：剪枝加低秩分解，再叠 Hadamard 旋转的 W8A8 INT8 量化，4 到 8 步不需要 CFG，9.1GB 权重让 24GB 消费卡跑 768p 带同步立体声的视频。权重沿用社区许可协议，欧盟、英国、韩国和美国未经授权不可使用。","要让一个视频生成模型跑进普通玩家的显卡，难点从来不是「能不能跑」，而是「跑得动不动、能不能用」。11 月 6 日，一家叫 Video Rebirth 的团队把 MiniMax H3 的 transformer 拆开重装，成果叫 HyperFlow H3 Flash：按其官方仓库给出的数字，约 310 亿参数的基座被压到 137 亿，生成视频的同时出同步立体声，量化后的 GGUF 权重最小一份 9.1GB，24GB 显存的消费级显卡就能跑 768p。这看起来是「开源视频模型继续下沉」的常规进展，但真正值得盯住的矛盾在另一头——这批权重继承了 MiniMax H3 社区许可协议，而那份协议把欧盟、英国、韩国和美国列为例外区域：真正想拿它在本地做创作的，恰好是被协议挡住的那些人。\n\n## 三刀：剪枝、低秩、旋转量化\n\nVideo Rebirth 在[仓库 README](https:\u002F\u002Fgithub.com\u002FVideo-Rebirth\u002Fhyperflow-h3-flash)里写明，他们对 MiniMax H3 的 transformer 只做了三处改动：\n\n- **结构化剪枝加低秩分解**：`fl2va` 分支砍掉 50 个 block 中的 3 个（block 6、21、22），102 个线性层做因子化；`ref2va` 分支只砍 1 个（block 9），因子化 114 个。两个分支分别落在 137 亿和 139 亿参数，模型卡标题写作 14B。\n- **旋转 W8A8 量化**：180 个线性层做 W8A8 INT8，前面加一层块对角 Hadamard 旋转，项目内部称之为 `convrot`。\n- **双时间条件 (t, r)**：每个 token 同时拿到当前噪声级 t 和步进终点 r，靠这个实现 4 \u002F 6 \u002F 8 步采样，而且不需要 CFG。\n\n## 9.1GB 权重能跑，但账单不止这些\n\nGGUF 档位从 Q3_K_M 的 7.2GB 到 Q8_0 的 14.9GB，Q4_K_M 是 9.1GB。运行时给了三条路：ComfyUI、Unsloth Studio 和 stable-diffusion.cpp，仓库里带一份 `hyperflow-two-time.patch`。\n\n官方在 1344×768、6 步、端到端（含文本编码与音视频 VAE 解码）下自测的 INT8 数字是：RTX 5090（32GB）出 5.2 秒片段要 92 秒、15 秒片段要 551 秒；RTX 4090（24GB）130 秒 \u002F 750 秒；RTX 3090（24GB）266 秒 \u002F 1684 秒。INT8 的峰值显存是 12.0–12.5 GiB（5.2 秒）和 21.8–22.4 GiB（15 秒）。同一张表里，数据中心的单张 H200 跑 5.2 秒片段是 36 秒、15 秒是 229 秒。\n\n有三个坑值得提前知道：\n\n- 16GB 系统内存是门槛，虽然模型本身只要 9.1GB；16GB 显存的卡官方明确还没测。\n- 参考输入非常贵：一张 2048 短边的参考图大约加进 14300 个 token，而片段本身 5.2 秒是 37710 token、15 秒是 109062 token，注意力开销随 token 数平方增长。两张图时，INT8 transformer 的耗时是纯文本生成的 2.6–2.9 倍，混着视频和音频的参考则是 5.2–5.8 倍。\n- 首次下载约 90GB（文本编码器换 fp8 是 68GB），其中 62GB 的基座 transformer 倒是不会下载。\n\n## 英伟达把它接进了 SOL-H3\n\nVideo Rebirth 说 SOL-H3 是他们的推理加速合作方，从第一天就支持这个模型。NVIDIA 那边也确实在 NVlabs\u002FSana 仓库里放了 `models\u002Fminimax_h3\u002FHyperFlow-Flash` 目录，并附了一份 VALIDATION.md。实测数据跑在 GB200 上，用的是 SOL-H3 自己的运行时而非上面那张自测表里的包，所以数字体系不同：测于 2026-11-05，checkpoint 锁在 `9e9ed14b`，分片与索引文件的 sha256 记在 PROVENANCE.json 里并逐一核对过；t2v 生成 5 秒片段，1 卡 13.0 秒、4 卡 3.4 秒、8 卡 1.9 秒；15 秒片段是 66.6 \u002F 16.1 \u002F 8.5 秒。4 卡 3.4 秒出一条 5 秒片子，确实到了实时级别。\n\n有意思的是那份 VALIDATION.md 自己列了未完成项：GPU 数值对齐门禁、1\u002F4\u002F8 卡冒烟测试、以及「用已发布的 remote code 重新计时」，三项都写着 in progress；t2v 也只列了 sol-engine 一个预设，官方预设那一栏因为用的是预发布版 INT8 内核而没有报告。厂商自己给的另一个口径是单张 B200 上相对 H3 基线提速 12.7 倍（Video Rebirth 官网数字）。\n\n## 权重跟着协议走\n\nMiniMax H3 社区许可协议（2026 年 8 月 2 日版本）里，「Excluded Territories」明确写着欧盟、英国、韩国和美国，适用地域是全球减去这四地。Video Rebirth 的权重被定义为 MiniMax H3 的 Model Derivative，按同一份协议分发，模型卡里还特意写明「This repository is not an official MiniMax product」。换句话说，未经 MiniMax 授权，这批权重在那四个地区不能使用。\n\n协议里另外两条也容易被忽略：商业产品年收入超过 2000 万美元需要单独书面授权；不得用 H3 及其输出去改进别的 AI 模型。管辖法域在香港。容易混淆的是推理代码本身是 Apache-2.0（ComfyUI 节点 GPL-3.0，stable-diffusion.cpp 补丁 MIT）——代码开源，不等于权重可以随便用。\n\n## 一点旁注\n\n旋转 INT8 这条路不是他们发明的，模型卡致谢里点名了 WAS\u002FMinimax-h3_Singularity_v2_pruned，称其在 MiniMax-H3 上的 rotated INT8 是 pioneering。HyperFlow 系列目前有三个成员：最早的 HyperFlow 降低训练成本，HyperFlow Infinity 拉长单次生成，H3 Flash 压延迟；Infinity 与最早的 HyperFlow 共用同一份权重，而针对全尺寸 MiniMax-H3 的 8 步 LoRA 则另开了一个仓库。\n\n## 所以呢\n\n社区把权重压下来的能力，已经快到像做家务一样；真正卡住分发节奏的，是那份协议还按国界画线。一个能在自己显卡上实时跑 768p 带声视频的模型，对欧洲和美国的开源爱好者来说依然只是纸面福利——你觉得这套「开放权重加区域许可」的组合，会先被改掉，还是会先变成常态？\n","https:\u002F\u002Fgithub.com\u002FVideo-Rebirth\u002Fhyperflow-h3-flash","998df6db-96e6-4b8e-8be1-cfa00a6cd177",[11,15,19,22],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":16,"name":17,"slug":17,"description":18,"color":14},"f3be854a-50f0-411c-893d-16d0df6def02","h3-series","MiniMax H3 专题：持续追踪 H3 的发布、开源、蒸馏与部署全链路",{"id":20,"name":21,"slug":21,"description":14,"color":14},"b49648f9-963e-4082-8684-3d085b7358fe","quantization",{"id":23,"name":24,"slug":24,"description":14,"color":14},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[26],{"id":27,"lang":28,"title":29,"summary":30,"content":31},"a563c3c9-a69b-4619-bcc2-f4b94224c08f","en","H3 squeezed to 13.7B: 768p video on a consumer GPU","Video Rebirth cut MiniMax-H3 to 13.7B with pruning, low-rank and rotated W8A8 INT8, so 9.1GB weights run 768p stereo video on a 24GB consumer GPU.","A third-party team has taken MiniMax's open-weight H3 video model and cut it down to fit on a consumer graphics card, while NVIDIA has already shipped a runtime for it. Here is what the compression actually consists of, what it costs to run, and why the license attached to these weights matters more than the speed numbers suggest.\n\n## Three surgical cuts\n\nVideo Rebirth states plainly in its repository that it changed the MiniMax-H3 transformer in only three ways.\n\nFirst, structural pruning plus low-rank factorization. The `fl2va` partition drops 3 of 50 blocks (blocks 6, 21 and 22) and factors 102 linear layers; the `ref2va` partition drops only 1 of 50 (block 9) while factoring 114. The two partitions land at 13.7B and 13.9B parameters respectively, against a base MiniMax-H3 transformer of roughly 31B, less than half.\n\nSecond, rotated W8A8 quantization. 180 linear layers are quantized to W8A8 INT8 behind a block-diagonal Hadamard rotation, which the project calls `convrot`.\n\nThird, two-time conditioning. Every token receives both the current noise level `t` and the step endpoint `r`, and that is what makes 4-, 6- or 8-step sampling work without classifier-free guidance.\n\nThe published repository is at https:\u002F\u002Fgithub.com\u002FVideo-Rebirth\u002Fhyperflow-h3-flash and the weights are on Hugging Face under VideoRebirthAI\u002Fhyperflow-h3-flash.\n\n## 9.1GB of weights, but the bill is longer\n\nGGUF quants run from Q3_K_M at 7.2GB up to Q8_0 at 14.9GB, with Q4_K_M at 9.1GB recommended as the default for 24GB GPUs. Three runtimes are supported: ComfyUI, Unsloth Studio and stable-diffusion.cpp, the last of which needs a `hyperflow-two-time.patch` applied.\n\nThe project's own end-to-end measurements are at 1344x768 with 6 steps, covering text encoding, the transformer and both the video and audio VAE decodes. On INT8, an RTX 5090 (32GB) takes 92 seconds for a 5.2 second clip and 551 seconds for a 15 second clip. An RTX 4090 (24GB) takes 130 and 750 seconds. An RTX 3090 (24GB) takes 266 and 1684 seconds. Peak VRAM for the INT8 path is 12.0-12.5 GiB at 5.2 seconds and 21.8-22.4 GiB at 15 seconds. In the same table, a single H200 takes 36 seconds for a 5.2 second clip and 229 seconds for a 15 second one.\n\nThree caveats are worth knowing before you plan around those numbers.\n\n- 16GB of system RAM is the floor, even though the weights themselves are only 9.1GB. Cards with 16GB of VRAM have not been measured yet.\n- Reference inputs are expensive. A single reference image at a 2048 short edge adds roughly 14,300 tokens, while the clip itself is 37,710 tokens at 5.2 seconds and 109,062 tokens at 15 seconds, and attention cost grows with the square of the token count. With two images, the INT8 transformer takes 2.6-2.9x its text-to-video time; with a mixed set of images, video and audio it takes 5.2-5.8x.\n- The first-run download is about 90GB, or about 68GB with the fp8 text encoder, though the 62GB base transformer itself is never fetched.\n\n## NVIDIA wired it into SOL-H3\n\nVideo Rebirth describes SOL-H3 as its inference-acceleration partner and says it supports this model from day zero. NVIDIA does ship it: the NVlabs\u002FSana repository contains a `models\u002Fminimax_h3\u002FHyperFlow-Flash` directory alongside a VALIDATION.md with full measurements.\n\nThose measurements were taken on GB200 using the Sol-H3 runtime rather than the package measured in the table above, so the two sets of numbers are not directly comparable. They were run on 2026-11-05, with the checkpoint pinned to revision `9e9ed14b` and byte-identical weight shards recorded in PROVENANCE.json. For text-to-video, a 5 second clip takes 13.0 seconds on one GPU, 3.4 seconds on four and 1.9 seconds on eight. A 15 second clip takes 66.6, 16.1 and 8.5 seconds respectively. Four GPUs producing a 5 second clip in 3.4 seconds is squarely into real-time territory.\n\nWhat is striking is that VALIDATION.md also lists what is not finished. The GPU numerics gate, the 1\u002F4\u002F8-GPU smoke tests and the re-timing with the released remote code are all marked in progress, and the text-to-video rows report only the `sol-engine` preset, with no `official` preset comparison because the earlier official cells used pre-release INT8 kernels and were therefore not reported. The provider itself estimates a 12.7x speedup over the H3 baseline on a single B200, a figure taken from Video Rebirth's own landing page rather than an independent measurement.\n\n## The weights carry the license with them\n\nThe MiniMax H3 Community License Agreement, dated August 2, 2026, defines \"Excluded Territories\" as the European Union, the United Kingdom, the Republic of Korea and the United States of America, and the \"Applicable Territory\" as everywhere else. Video Rebirth's weights are distributed as a Model Derivative of MiniMax-H3 under that same agreement, and the model card states explicitly that the repository is not an official MiniMax product. In practical terms, without authorization from MiniMax, these weights cannot be used in those four regions.\n\nTwo further clauses tend to get overlooked: commercial products and services generating more than 20 million US dollars in yearly revenue need separate prior written authorization, and the works and their outputs may not be used to improve any other AI model. Jurisdiction and governing law sit in Hong Kong.\n\nIt is easy to conflate the code with the weights here. The inference code on GitHub is Apache-2.0, with the ComfyUI custom nodes under GPL-3.0 and the stable-diffusion.cpp patch under MIT. Open code does not make the weights free to use.\n\n## A side note on where the idea came from\n\nThe rotated INT8 route was not invented here. The model card credits WAS\u002FMinimax-h3_Singularity_v2_pruned with pioneering rotated INT8 on MiniMax-H3. The HyperFlow series now has three members: the original HyperFlow, which reduced training cost, HyperFlow Infinity, which extended the length of a single generation, and HyperFlow H3 Flash, which cuts latency. Infinity shares its weights with the original HyperFlow, while the 8-step LoRA for the full-size MiniMax-H3 transformer lives in a separate repository.\n\n## So what\n\nThe community's ability to compress weights has become close to routine. What is still throttling distribution is a license that draws lines along national borders. A model that can generate 768p video with synchronized stereo audio in real time on a consumer GPU remains, for open source enthusiasts in Europe and the United States, a paper benefit rather than a working one. Do you think this combination of open weights with territorial licensing gets fixed first, or becomes the norm?\n","hyperflow-h3-flash-int8-consumer-gpu-video","2026-11-06T08:30:00Z","2026-11-08T01:11:36.280317Z","2026-11-08T01:11:36.280331Z",true,"agent",683,[40,49],{"slug":41,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":36,"created_at":47,"modified_at":48},"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":17,"tag_slug":17,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":36,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"a2e06c5f-ab65-468f-b936-c654ff75aa38","14 秒视频 9 秒出片:Video DeltaNet 给 H3 换混合注意力","video-deltanet-h3-hybrid-attention","2026-09-18T23:08:51+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"02c8b500-ec11-44a6-8c58-6e880563dad8","FastH3 开源:4 步蒸馏版 MiniMax H3,B200 单卡最高提速 14 倍","fasth3-4-step-distilled-minimax-h3","2026-08-30T21:30:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"06b504ab-f9ac-46b2-a1e0-acfbe0fd3996","上交开源WaveForcing:流式生视频134.8帧\u002F秒","wave-forcing-streaming-video-generation","2026-11-03T13:05:45+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"bf2c4f0f-3911-4227-b741-7252b1875cc7","FrameForge:视频模型跨行图生图,帧距自己挑","frameforge-adaptive-frame-gap-image-generation","2026-11-02T17:13:17+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"85b85139-0d39-464a-99cf-415fd498e9d2","L3G-World:高斯记忆让世界模型9 FPS实时生长","l3g-world-gaussian-memory-realtime","2026-10-30T21:12:08+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"703ab8f9-4f6c-46fd-8b3c-f53387619cad","更稀疏、更清晰:视频扩散注意力砍掉九成","ab-attention-block-sparse-video-diffusion","2026-10-30T15:15:10+00:00"]