MiniMax H3 Lands Day-0 in ComfyUI: First Open-Weights Video Model With Native Stereo Audio and 2K Output
ComfyUI Blog / Hacker News·high signal
Released August 3 (309 points, 87 comments), MiniMax H3 is the company's first video model shipped with open weights, accepting text, image, video, or audio input and producing up to 2K resolution, 15-second clips with stereo audio generated simultaneously rather than dubbed afterward. Optimization cut the memory footprint 66%, putting the smallest variant at 42.5 GB and making it runnable on an RTX 3060 with dynamic VRAM offloading. Weights are on Hugging Face as Comfy-Org/MiniMax-H3, requiring ComfyUI 0.30.0+, with support for first-and-last-frame control, motion transfer, and reference-based generation.