A single sentence posted by Ethereum co-founder Vitalik Buterin on August 7 (22:03 KST, or 13:03 UTC) shook the open video model camp. His assessment: "the first open video model I've seen that surpasses HunyuanVideo 1.5."
A Long-Held Throne Shaken for the First Time
Buterin noted that it is "an impressive achievement in itself" that Tencent's HunyuanVideo 1.5 held the throne for so long, but said MiniMax H3 is the first open-weights model to surpass it. His remark carries weight because he has lamented the drift of the image and video frontier toward closed models.
The Skill in Numbers
MiniMax H3 is a 33-billion-parameter omni-modal generation model released on July 31. It takes text, images, video, and audio as a single input and generates 4 to 15 second videos at up to 2K resolution and 24 fps, along with native stereo audio that includes dialogue, sound effects, and room tone. On the independent Artificial Analysis benchmark, it ranked 1st in video editing, 2nd in text-to-video, and 3rd in image-to-video.
"Open" Has Exceptions: The US, EU, UK, and South Korea
The weights were released on August 3 on Hugging Face (MiniMaxAI/MiniMax-H3) at 42.5GB. However, the MiniMax Community License excludes the US, EU, UK, and South Korea from its "covered territory". That means in these regions, no license is granted to run, modify, or distribute locally downloaded weights or to use their outputs. It is a reason to read the license text before taking the word "open" at face value.
Judge Performance and License Separately
The performance assessment is grounded: it is Vitalik's personal comment and is also backed by an independent benchmark. But commercial and legal judgments are a separate matter. If you plan to actually run this model or use its outputs in the US, EU, UK, or South Korea, it is safest not to skip rereading the license terms in the original text.
Sources: vitalik.eth's X post · Tech Times report