MiniMax open-sources H3, a 33-billion-parameter omni-modal video model, with day-0 ComfyUI support

MiniMax released open weights for H3 on August 3, 2026, a 33-billion-parameter omni-modal generative model that jointly processes text, images, video and audio and generates video with native stereo sound.

The model outputs clips at resolutions up to 2K, durations up to 15 seconds and 24 frames per second, across aspect ratios from 21:9 to 9:16. The open release ships two task-specific checkpoints: FL2VA, for text-to-video with first- and last-frame control, and Ref2VA, for reference-based generation from an image, video or audio input.

What makes the launch usable rather than just downloadable is tooling. ComfyUI, the open-source node-based interface for generative models, shipped native support the same day, adding three official workflow templates for text-to-video, image-to-video and reference-based generation. That kind of day-zero integration is unusual and lets people run H3 through an existing pipeline immediately.

MiniMax also cut the model’s memory footprint sharply. About 13 billion of the 33 billion parameters sit in branches that can be cached for inference-only use, and the smallest variant trims memory from 123.6 GB in full precision to 42.5 GB, a 66% reduction. MiniMax and ComfyUI say that is low enough to run the 2K video model locally on a consumer graphics card such as an RTX 3060.

Those numbers come from the two companies releasing the model and its tooling, not from independent testing, and real-world speed and quality on a mid-range GPU remain to be measured. Still, an open-weight video model that fits on widely owned hardware lowers the bar for local generation well below the cloud services that have dominated the category.