You can now run MiniMax H3’s 2K video model on a consumer GPU

You can now generate 2K video with native stereo sound on consumer hardware.

MiniMax released open weights for H3, its third-generation video model following Hailuo 01 and 02. ComfyUI added day-zero support for it today in version 0.30.0 and on Comfy Cloud. According to ComfyUI's release post, the model takes text, images, video, or audio and generates up to 15 seconds of 2K video, with native stereo audio created in the exact same pass.

Why it matters: Most AI video workflows require chaining separate tools for motion and sound. H3 handles both in one step, doing text-to-video, image-to-video, and reference-based motion transfer. Heavy optimization—including weight pruning and int8 quantization—slashed memory requirements by 66%, shrinking full precision from 123.6 GB down to 42.5 GB so it can run locally on GPUs like an RTX 3060.

Try it: Update ComfyUI to v0.30.0, grab the model weights from Hugging Face (Comfy-Org/MiniMax-H3), and download the starter templates directly from the ComfyUI workflow library.

High-end local video generation just lost a massive hardware tax.