News · 2026-08-16

MiniMax put its video-with-sound model on Hugging Face

MiniMax has released the weights for H3, its multimodal generation model that produces video with sound built in rather than added afterwards. The model landed on Hugging Face on August 13, 2026, about two weeks after the Chinese lab said at launch that it planned to open the weights, and it has since been downloaded more than 2.3 million times.

Key facts

• H3 generates video with native 32 kHz stereo audio, 4 to 15 seconds long at 24 frames per second, with 2K resolution available through a separate regeneration step.
• The weights reached Hugging Face on August 13, 2026, following the July 31 launch announcement.
• The model page shows more than 2.3 million downloads and over 4,000 likes.
• Primary sources: the Hugging Face model card and MiniMax's launch post.

At launch MiniMax described H3 as "a general-purpose multimodal generation model" that "understands unified context across text, images, video, and audio", and committed to opening it: "we plan to open up the model weights in the coming days, subject to applicable laws and regulations." The weights arrived roughly two weeks later.

The technically interesting part is the unified context. Most video generation tools take a prompt and maybe a starting image. H3 takes a mixed bag and works out the relationship between the pieces. MiniMax's own example prompt is instructive: "Reference the Hitchcock camera movement from Video 1, have the character in Image 2 sing, with the vocals matching Audio 3." You describe how the inputs relate to the output you want, in words, and the model handles the rest. That is a different interaction model from stringing together a video generator, a voice model, and an editor.

Audio is the other departure. Video models have generally treated sound as a post-processing problem, generating the picture and then finding something to lay under it, which is why AI video so often sounds subtly detached from what is on screen. H3 produces 32 kHz stereo alongside the frames, with stable dialogue support in 11 languages including Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.

On specifications, the model card is precise: output runs 4 to 15 seconds at 24 frames per second, with a default short side of 768 pixels and 2K reachable through a separate regeneration pass, across aspect ratios from 21:9 through 9:16. MiniMax positions it for commercial content work, claiming strength in instruction following, accurate text and brand rendering, and video-to-video motion transfer, and says that at 2K its per-second price is less than a third of mainstream competitors.

The reason a weight release matters more than another hosted video tool is control. A hosted video model is a rate limit and a content policy you do not set. Weights on disk can be fine-tuned on a specific visual style, run on hardware you already own, and integrated into a pipeline that does not phone home. MiniMax says hardware compatibility was a design consideration from the earliest stages, and the reasoning it gives for opening the model is that closed video generation has produced slower iteration and a less open ecosystem than the language model field.

The honest caveat is the licence, and it is the same one Ground Truth flagged when MiniMax released its five-minute song model the following day. H3 ships under a MiniMax H3 Community License Agreement, not Apache 2.0 or MIT. Open weights and open licence are different things, and a company planning to build on this needs to read the terms rather than assume. The performance claims in the launch post are also the vendor's own, with no independent evaluation yet, and the 2K path runs through a separate regeneration model rather than being the base output.

Practically, there are three ways in. The weights are on Hugging Face, the code and official prompt-writing guidance are in the GitHub repository, and the hosted version runs at hailuoai.video for anyone who wants to try it before downloading tens of gigabytes.

The wider pattern is now hard to miss. Chinese labs have spent this year putting frontier-adjacent multimodal weights into public hands while American labs largely have not, and generative video, one of the last areas where closed models had a clear lead, is the newest front. Background reading: diffusion models, open-weight models, and neural text-to-speech.

Primary source, verified: read the paper →

Ground Truth. (2026, August 16). MiniMax put its video-with-sound model on Hugging Face. Ground Truth. https://groundtruth.day/news/minimax-put-its-video-with-sound-model-on-hugging-face.html

@misc{groundtruth:minimax-put-its-video-with-sound-model-on-hugging-face,
  title  = {MiniMax put its video-with-sound model on Hugging Face},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/minimax-put-its-video-with-sound-model-on-hugging-face.html}
}

Topics: video-generation · open-weights · minimax · china · multimodal · audio