MiniMax has released the weights for H3, its 33 billion parameter video generation model. It’s the first open-weight model to top a public video generation ranking: Artificial Analysis places it first in Video Editing, second in Text-to-Video, third in Image-to-Video.

Until now, convergence between open and proprietary models concerned text: DeepSeek V4-Flash matched GPT-5.6 Luna at 60% lower cost. Video remained the territory of labs with massive compute and closed models. H3 brings open-weight video generation to measurable quality, doing so with a model that processes text, images, video and audio together, generating clips from 4 to 15 seconds with native stereo audio.

For content creators, the concrete benefit is fine-tuning. You can train the model on your footage, your characters, your visual style. With a closed API you can’t.

Two pieces remain closed. The 2K resolution module and H3-Context-IR, the system that translates complex prompts into structured format before generation, are not in the weights. Anyone running it locally on ComfyUI tops out at 768p and must handle context preparation themselves, following MiniMax’s guides. The license permits commercial use only below $20 million in annual revenue.

That same day ByteDance released Seedance 2.5, a closed model that generates 30-second clips with integrated audio. Competition in generative video hasn’t stopped waiting for H3.

In detail

Until H3, public video generation rankings were dominated by proprietary models. Measurable quality on benchmarks like VQEG and VMAF was the domain of labs that don’t release weights. Open-weight video existed (SANA-Video from NVlabs, for example), but hadn’t yet topped a public leaderboard.

H3 is an omni-modal 33 billion parameter model. It takes as input a combination of text, images, video and audio, and outputs a video clip with stereo audio at 32 kHz. A single prompt can include up to 9 reference images, 3 video clips and 3 audio clips, for a maximum of 12 total files. Output duration ranges from 4 to 15 seconds, at 24 frames per second.

The complete system has three modules. H3-Context-IR is the preprocessor: it takes complex prompts and reference material, interprets them, and converts them into a structured intermediate representation the generation model can use. MiniMax describes it as critical to output quality, but doesn’t include it in the open weights because it depends on a pipeline of hosted models and services. It provides an API to reproduce its behavior and detailed guides to build your own.

H3-Base is the actual generative model. From open weights it produces 768p resolution output. This is what you download from Hugging Face and can run locally via Diffusers or ComfyUI.

H3-Regenerate-2K is the upscaling module that brings output to 2K by leveraging both H3’s generative capabilities and information from the original context. This also remains closed.

Anyone downloading the weights today gets H3-Base. To achieve results comparable to those in the rankings, they must build their own context preparation pipeline following MiniMax’s published guides, or call the H3-Context-IR API as a service.

The community license permits commercial use for companies under $20 million in revenue. Above that threshold a separate commercial agreement is needed. It’s a restriction already seen in Chinese open models: weights available, but with a revenue cap that excludes large companies.

On the competition front, ByteDance released Seedance 2.5 the same day. Closed model, 30-second clips with integrated audio. The message is clear: generative video is a front where competition runs hard, and those who open their weights do so strategically, not generously.