MiniMax 发布 H3 并开放权重，但“全面开源”和 SOTA 说法需谨慎

Artificial intelligence Tech # Video AI: MiniMax challenges ByteDance with low price, open weights for new H3 model Firm aims to ‘support open-source community’ and a ‘broad range of hardware’, as H3 model bests Seedance 2.0 in some tasks 2-MIN READ2-MIN Listen People walk past the MiniMax booth during the World Artificial Intelligence Conference in Shanghai on July 17. Photo: Reuters Xinmei Shen Chinese AI firm MiniMax has launched H3, its newest multimodal video generation model, pledging to break closed-source “dominance” through open weights and competitive pricing – as rival ByteDance rolls out its latest Seedance 2.5 model. [...] H3 was currently the world’s most powerful AI model in video editing, according to benchmark platform Artificial Analysis. However, it trailed Google’s Gemini Omni Flash in text-to-video tasks, and ranked behind both ByteDance’s Seedance 2.0 and Gemini Omni Flash in image-to-video capabilities. Designed for commercial content creation across fil

To reduce the computational cost of long multimodal sequences, H3 natively supports sparse-attention training and inference. The initial open-source release provides inference with full attention only. Our sparse-attention implementation will be released in a future update. [...] The model uses three-dimensional Multimodal Rotary Position Embeddings (MM-RoPE) to represent positional relationships across the temporal and two spatial dimensions, `(t, h, w)`. During the final stage of training, we introduce native sparse attention to reduce the computational cost of long sequences. The sparse-attention implementation is not included in the initial open-source release and will be published separately in a future update. [...] Due to the complexity of the system, this module is not yet open-sourced. We will release it once it is ready. We provide an API for validating the official results; see "Full 2K Workflow" below.

只要涉及到一切跟视觉包装相关的工作，它干的都极佳。 如果说，Seedance更偏向与前期，更重影视和特效，那H3我觉得，就更偏向后期，更重广告与动效。 两者相辅相成，终于补全了整个视频制作的全流程。 网址在此： 在首页上，你就能看到这次全新的H3了，也是一个原生多模态模型。 在纯粹的功能上，非常的全面，全能参考啥的也都有。 支持视频图片音频等等的各种参考，最多12个。 参数上，目前开放的直接是2K，甚至768P的都暂时还没开放，最多15s。 价格上2K的效果，1秒是12贝壳，差不多就是10块钱一条15秒的2K视频。 动作参考、指定元素替换、换背景、改写视频、实拍视频加特效、绿幕换背景、白模视频上色、视频续写延长、视频整体参考等等这些常用的功能，也都没啥问题，而且效果不错。 这次的重点，我们自然还是要回到，视觉包装上。 其实我们会发现在过去半年里面，AI视频加速的最多的是AI短剧赛道。 而传统的商业广告、宣传片、综艺等等，AI介入的好像还没有那么多。 原因特别简单，过去我们一直在做的AI视频模型更像是一台很强的虚拟摄影机，它们能够凭空创造一个世界，然后把这个世界拍下来，记录下来。 但是，摄像机拍完以后，一条真正意义上的视频，其实才刚刚开始。 因为人类最终看到的绝大多数视频，都不是现实世界原封不动的切片。 我们会剪辑，会加字幕，会设计片头，会制造转场，会用图形、文字、UI和特效重新组织信息。 这些东西，现实世界里原本都不存在。 它们全部是人类创造出来的视觉语言。 这些就是视觉包装。 最典型的就是综艺里面经常会加的花字。 这种以前AI视频几乎没有办法去处理，因为很难生成稳定加准确的文字，且很难理解文字与视频素材和视频内容之间的关系。 但是这一次，MiniMax H3可以搞定这种了。 [...] # MiniMax H3发布，还要开源，AI视频终于有了自己的后期之王 头像 2026-07-31 10:21发布于北京科技领域创作者 问AI · AI视频模型为何长期忽视视觉包装需求？ 今天，被Seedance 2.0压制了整整半年，沉寂了很久的AI视频模型行业，终于又有了点新的动向了。 MiniMax，终于掏出了他们的H3模型。 最后绕不开的那个名字，依然是Seedance 2.0。 它就像那一堵最高的山。 以至于过去

| License | MiniMax H3 Community License AgreementSource-available, with territory restrictions — check the repo before deploying | [...] ## Introducing MiniMax H3 Video generation has been the last major modality where the closed labs kept a comfortable moat. Weights stayed private, iteration was slow, and if you wanted frontier-quality video you rented it through someone else's API. That changed twice in the space of three days. MiniMax announced H3 on July 31, 2026 as an API-only product, then published the weights to Hugging Face on August 3 under the MiniMax H3 Community License. It's the strongest open-weight video model released to date, and unlike most of the "open" video releases before it, the audio isn't a separate stage bolted on afterward. Here's what H3 actually is, what MiniMax did and didn't open-source, and how to size GPUs for it on Runpod. ## What makes H3 different [...] ## A note on the license We would be remiss if we did not mention that the license appears

Hailuo H3 Text-to-Video (T2V): 👉 Hailuo H3 All-In-One Reference Video Gen (Supports 9 Images, 3 Videos & 3 Audios): 👉 #aigc #comfyui #ai #runninghub #workflow #huggingface #github #aivideogenerator 2 comments ### Transcript: [...] # 彻底变天！MiniMax海螺H3官宣开源，硬刚Seedance2.0！ / MiniMax Conch H3 Goes Open Source to Challenge Seedance 2.0! ## AIGC-Singularity 1710 subscribers 7 likes ### Description 241 views Posted: 2 Aug 2026 #hailuoai #minimax #aivideo #aitutorial #seedance2 🔔 关注我们 (Stay Connected) X (Twitter)： 欢迎关注 X，后续新消息将第一时间在此发布。 Civitai： 🌍 English Version for Overseas Users [Limited Time Offer: Free Online RTX 4090 Access!] Experience various fun and interesting AI applications with powerful RTX 4090 graphics cards online for free! Runninghub Link: Registration Bonus: Register through the link above to get 1000 RH coins instantly! 🎬 Hailuo H3 Workflows (RunningHub): Hailuo H3 Image-to-Video (I2V & Start/End Frame Control): 👉 Hailuo H3 Text-to-

# 开源的春天，闭源的噩梦！MiniMax-H3效果炸裂，seedance不再一家独大！ ## yue gu 559 subscribers 30 likes ### Description 1125 views Posted: 3 Aug 2026 #runninghub #comfyui #海螺ai 👇 点击下方链接注册，领取1000 RH币免费在RTX4090上运行： 国内 (.cn)： 全球 (.ai)： 国内工作流： 海外工作流： 夸克网盘： 文档： 1 comments ### Transcript: