MiniMax has released H3, a 33B-parameter video generation model that produces video with synchronized stereo audio and supports multiple input modalities: text, images, video, and audio references. The model handles text-to-video, first/last frame control, motion transfer, scene editing, and clips up to 15 seconds at ~768p resolution. Under the hood, H3 uses a dense Omni-Transformer paired with Qwen3-VL-32B as its encoder.

ComfyUI has already shipped optimized builds with INT8, FP8, and NVFP4 quantization, bringing minimum download size down to roughly 42.5 GB — runnable on consumer GPUs with RAM/SSD offloading. However, calling H3 “open source” requires a caveat: only the base weights are published. The H3-Context-IR system for complex multimodal queries remains server-side, 2K regen is API-only, and sparse attention is promised “later.” A significant license restriction blocks public use of the weights in the EU, US, UK, and South Korea without separate permission from MiniMax.

Related: MiniMax M3 Open-Weight Release, MiniMax M3 Pro Details

Model on Hugging Face · ComfyUI Build