MiniMax H3

MiniMax H3 is an open multimodal model from MiniMax that creates video with native stereo sound (32 kHz). Clips run from 4 to 15 seconds, the base resolution is 768p, and 2K is produced by a separate H3-Regenerate-2K module. Modes include text-to-video, image-to-video (including first and last frame) and omni-reference mode with up to 9 images, 3 videos and 3 audio clips. The model voices dialogue in 11 languages, including Russian. Inside are a 33-billion-parameter transformer and an encoder based on Qwen3-VL-32B. The prompt-preparation module H3-Context-IR is not part of the open release. The weights are published on Hugging Face under the MiniMax H3 Community License Agreement. Per the license text, there are territorial restrictions, check the license, and products with yearly revenue above 20 million US dollars need written permission from MiniMax; other terms are in the license file in the model repository. The author does not state hardware requirements; the deployment example uses 4 GPUs. You can try it without installation in the Hailuo AI app from the same developer, which is in our catalog: Hailuo AI.