MiniMax just posted the weights for its H3 video-generation model on Hugging Face, so anyone with a decent GPU can download the core system. The release stops short of the full commercial pipeline – only the H3-Base generator is open-weight; the preprocessing and upscaling stages stay behind an API.

What the open-weight release actually contains

MiniMax’s H3 architecture splits into three logical blocks. The two that remain API-only are:

  • Context-IR layer – preprocesses input text, images or video snippets.
  • Regenerate-2K layer – boosts the output to high-resolution (2 K) video.

Developers can pull down the H3-Base core generator, the engine that turns multimodal prompts into a 768-pixel-short-edge video clip with native audio.

Hardware and runtime realities

The model contains 33 billion parameters and ships in two flavors:

  • fl2va – accepts text and image prompts.
  • ref2va – accepts a reference video as part of the prompt.

The smallest quantized checkpoint is about 21 GB; the full-precision checkpoint is 123.6 GB. MiniMax recommends at least 42.5 GB of VRAM or storage for comfortable operation.

On a 12 GB GPU, developers can run the model with CPU offloading, but inference slows noticeably. Local generation tops out at 768 px on the short edge, far below the 2 K resolution the proprietary Regenerate-2K layer delivers.

Why the release matters

Artificial Analysis’s recent ranking puts H3 first in the “video editing” category and in the top three for both “text-to-video” and “image-to-video”. Those scores make it the most capable open-weight video model available today.

The model is multimodal: it understands text, images, video, and audio, and it can synthesize video with synchronized sound—an uncommon breadth for a publicly downloadable weight set.

Limitations and the other side of the story

H3 does not win everywhere; Gemini Omni Flash outperforms it in certain tasks. Because high-resolution upscaling lives behind an API, developers cannot reach the full commercial output without paying for MiniMax’s service.

Licensing adds another hurdle. The terms explicitly ban use in the European Union, United Kingdom, South Korea, and the United States. Organizations in those regions must negotiate a separate commercial agreement or look elsewhere.

Takeaway: MiniMax’s open-weight H3 gives developers a powerful multimodal video generator they can run locally, but the API-only preprocessing and upscaling, hefty hardware needs, and regional licensing bans keep the full commercial experience out of reach for many.