MiniMax H3 (Hailuo 3)

Open-weight omni-modal video model from MiniMax (Hailuo 3): text, image, video and audio in, video out with natively synchronised stereo audio. The catalogue's first video model that generates its own soundtrack rather than needing a separate TTS or music pass. Note the open weights are H3-Base, validated at a 768-px short edge; the 2K stage (H3-Regenerate-2K) has NOT been open-sourced, so there is no local 2K path. LICENCE - READ BEFORE DOWNLOADING. The MiniMax H3 Community License Agreement grants rights only inside its "Applicable Territory", defined as worldwide EXCLUDING the European Union, the United Kingdom, the Republic of Korea and the United States of America. The restriction extends to Outputs, i.e. generated video, not just the weights. Readers in those territories can apply for a per-deployment licence at platform.minimax.io/h3-license. The HuggingFace metadata names the agreement (license_name / license_link) but the licence TAG is only "other", which carries no territorial signal - the restriction is visible solely by reading the LICENSE file. Nor does anything downstream relax it: the Comfy-Org repack carries the same license_name, and the Apache-2.0 Turbo LoRA is an adapter on H3, whose Model Derivatives clause (iii) expressly covers distillation. Hardware: do not size this from the 498 GB repo, which is bf16 masters plus duplicate stacks. The ComfyUI working set is four files totalling 42.5 GB, and ComfyUI evicts modules on demand, so the binding term is the largest single module - 19.53 GiB. Smaller cards run it via dynamic VRAM offload, paying in time rather than failing: Comfy-Org's launch post names the RTX 3060 12 GB, and system RAM becomes the real constraint.