MiniMax H3 open weights: what you can actually run locally
In early August, MiniMax published the weights of its H3 video model on Hugging Face. Clips with stereo sound can now be generated on your own machine — no subscription, no per-generation API fees. But before downloading hundreds of gigabytes, keep two things in mind: the full 2K pipeline did not make it into the open release, and the license does not permit everyone, everywhere.
What exactly was released
The MiniMax-H3 repository ships two task-specific checkpoints. FL2VA creates video with sound from text and can animate frames: you provide the first, the last, or both. Ref2VA builds a clip from references — a single request can include up to nine images, three videos and three audio files, so a character's look, camera movement, style and voice are each set by their own example. The core is a dense 33-billion-parameter transformer, the weights are published in BF16, and the whole repository weighs around 288 GB.
What the model can do
H3 takes text, images, video and audio as input and produces clips of 4 to 15 seconds at 24 fps — with built-in stereo audio that the model synthesizes itself, in sync with the picture. Base resolution is 768p on the short side. Audio is not a separate processing stage here but part of the same model: you can set a voice or musical rhythm by reference, which changes the usual workflow noticeably.
About 2K — an important caveat
The advertised 2K does not come from the open part of the pipeline. A separate module called Regenerate-2K handles the resolution boost, and MiniMax did not release it: the official route to 2K is to send your local 768p result back to the company's cloud API. Context-IR, the module that parses complex combinations of prompts and references, is not open either. So "running everything locally" currently means "running base generation".
Will an ordinary computer cope?
The ComfyUI team added support on day zero and heavily optimized the run: modulation weights — about 40% of all parameters — were replaced with a compact lookup table, and int8 quantization shrank the total footprint from 123.6 GB to 42.5 GB. Combined with dynamic offloading to system RAM, that is enough for the model to run on a card as modest as the RTX 3060. The community went further: WanGP configurations for cards with 5–6 GB of VRAM have appeared, using aggressive RAM offloading, reduced resolution and long minutes per short clip. Keep in mind these are community recipes, not official requirements: detail comes out softer, and stability depends on your RAM size and disk speed. For full quality, the developers recommend top-tier cards like the RTX 5090.
Licensing: this is not open source in the usual sense
The weights are open but distributed under a Community License with conditions. Commercial use is allowed for companies with under $20 million in annual revenue — larger businesses need a separate agreement. And crucially, geography: the license only covers the US, the European Union, the United Kingdom and South Korea; elsewhere use requires written permission from MiniMax. "Open weights" is the accurate term, not "fully open source".
Who really needs a local run
A self-hosted instance makes sense when privacy, high generation volume or fine-tuning on your own data matters — exactly what open weights enable. If you just want a quality clip fast, without a GPU farm or environment hassle, the cloud is easier: in the NeuralSpace video section clips take a couple of clicks, and we covered what H3 itself can do in our separate review.
FAQ
What GPU does MiniMax H3 need?
For comfortable work, top-tier cards with plenty of memory. In practice people run base generation even on 6 GB of VRAM through ComfyUI and WanGP, trading away resolution and speed: a short clip can take 10–15 minutes.
Can I use H3 commercially?
Yes, if your company's annual revenue is under $20 million — that is a Community License condition. Larger organizations need a separate agreement with MiniMax.
Why isn't H3 fully open source?
Only the base part of the pipeline is open: the modules for upscaling to 2K and parsing complex references remain closed, and the license restricts commercial use and regions.