MiniMax released H3, its first open-weight video model, and it collapses most of a video pipeline into a single checkpoint. One model handles text-to-video, image-to-video, first-and-last-frame control, reference-to-video driven by images, clips, or audio, and motion-transfer editing that lets you revise an existing shot instead of regenerating it. Output runs at 24 FPS up to 2K, up to fifteen seconds a clip, with real stereo audio generated jointly with the picture rather than dubbed on afterward. The weights went up on Hugging Face with day-0 support in ComfyUI and vLLM, and the local-inference story is the surprise. ComfyUI found that the model's modulation weights — roughly 40% of its parameters — could be pruned and replaced with a functionally equivalent lookup table. That, together with int8 convolution quantization and custom kernels, cut the smallest variant's footprint from 123.6 GB to 42.5 GB. With dynamic VRAM offloading it runs on an RTX 3060, and comfortably on a 5090. Community clips of fantasy battles and creature animation went up within hours. The catch is legal, not technical. The H3 community license grants rights worldwide excluding the European Union, the United Kingdom, the Republic of Korea, and the United States — the four jurisdictions where generative-video copyright litigation and AI regulation are moving fastest, and where MiniMax is itself a defendant in the Disney, NBCUniversal, and Warner suit over Hailuo. Inside the licensed territory it's free for non-commercial use and for companies under $20M in revenue with attribution. Everyone else — including U.S. creators — can apply to MiniMax for formal authorization, or use the hosted API, which remains globally available.