MiniMax H3 pushes open AI video toward production
August 2, 2026
MiniMax has introduced H3: a video model for 2K clips with native stereo audio and announced open weights. The key question is whether the weights actually arrive.
What this is about
MiniMax introduced MiniMax H3 on July 31, 2026. The model generates videos up to 15 seconds long, at 2K resolution, with native stereo audio. The real hook is not just resolution: MiniMax says it plans to release open weights, moving H3 closer to the open model world that is far more common in language models than in video models.
That matters because AI video is still dominated by closed platforms. If teams cannot run or adapt models themselves, they remain dependent on the pricing, limits, moderation, and product choices of individual providers.
What MiniMax H3 actually does
According to MiniMax, H3 understands text, images, video, and audio together as context. A user can combine several references: a camera move from one video, a character from an image, and a voice from an audio file. The model is supposed to turn those into one coherent clip instead of treating each modality as a separate task.
Independent summaries confirm the core claims: up to 15 seconds, 2K output, 24 frames per second, and audio generated together with the visuals. MiniMax also names use cases including advertising, e-commerce, product design, UI/UX, and gaming.
Why it matters
Open weights would change the market if MiniMax actually ships them. Studios, developers, and research teams could test the model locally, adapt it to their own hardware, and build workflows that are not fully tied to one cloud. That matters because video generation is expensive and professional teams need reproducible results.
The price is also a signal. Reports cite 0.8 yuan per second for 2K generation, or 12 yuan for 15 seconds. If that level remains stable, more expensive closed video APIs will face pressure.
In plain language
Imagine a film kitchen. Until now, picture, sound, and editing often had to be cooked in separate pots, with everyone hoping the result would fit together. H3 promises one shared pot: image, voice, motion, and style are planned together. Open weights would be the recipe, not just the finished meal from a restaurant.
A practical example
A small furniture studio wants to create 40 product clips per month. Each clip shows a shelf in an apartment, with a camera move, room sound, and a short usage scene. At 15 seconds per clip, that is 600 seconds of material per month.
If H3 stays cheap enough through the API, the studio can test many variants. If the weights later become genuinely available, a service provider could adapt the model to the studio's own product photos, materials, and visual language. That is where the production value sits: not in one clip, but in repeatable campaigns.
Scope and limits
- The weights have been announced, but checks on August 1, 2026 found that they were not yet publicly downloadable.
- Open weights do not automatically mean open source; training data, code, and license details may remain restricted.
- A model that can imitate brand styles from references can also make misuse easier, including fake product videos or misleading advertising.
SEO & GEO keywords
MiniMax H3, Hailuo 3.0, AI video, open weights, generative video models, 2K video, native audio, video-to-video, China AI, creative tools
💡 In plain English
MiniMax H3 makes AI video more interesting because visuals, sound, and multiple references are processed together. The real test is not the demo, but whether the announced open weights actually become usable.
Key Takeaways
- →MiniMax introduced H3 on July 31, 2026.
- →The model generates up to 15 seconds of 2K video with native stereo audio.
- →MiniMax has announced open weights, which were not publicly available at the time of checking.
- →The reported price of 0.8 yuan per second could pressure closed video APIs.
- →Reference control is useful for production, but also increases misuse risks.
FAQ
Are the MiniMax H3 weights already available?
At the time of checking on August 1, 2026, independent reporting said they were not yet publicly downloadable. MiniMax has announced a release.
What makes H3 different from many video tools?
H3 is designed to understand text, images, video, and audio together, then generate a clip with native stereo audio.
Why does this matter for creative teams?
If the price and weights hold, teams could build cheaper and more controllable production workflows instead of relying only on closed web tools.