MiniMax released H3 on July 31, a multimodal generation model that produces 2K video with stereo audio in a single forward pass. The company published the weights to Hugging Face on August 3 under the MiniMax H3 Community License, making it the most capable open-weights video model currently available.
H3 is not a text-to-video model with audio bolted on afterward. It is a single transformer that accepts text, images, video, and audio as unified input context and generates 4-to-15-second clips with native stereo sound. A user can drop a product photo, a motion reference clip, and a voice recording into one prompt, describe the relationship between them, and H3 resolves the cross-modal work itself. The model supports up to 9 reference images, 3 video clips, and 3 audio tracks per generation.
The architectural change that makes this practical is H3-VAE, a rewritten tokenizer with a compression ratio that MiniMax claims yields roughly a 4x gain in effective sequence length. This is what makes native 2K output economically viable. At 2K, MiniMax says per-second pricing is less than a third of competing models; at 768p, less than half the price of comparable 720p offerings.
ComfyUI shipped day-zero support. The model runs locally on an RTX 3060, which puts open-weights video generation with native audio on consumer hardware for the first time. RunPod published a guide the same day on GPU sizing for self-hosted inference.
On the Artificial Analysis video leaderboard, H3 ranks first in video editing but trails competitors in text-to-video and image-to-video. The open weights may shift those rankings as the community fine-tunes.
MiniMax is billing this as a commercial tool for advertising, branding, e-commerce, and game cinematics. The model is the third generation in the Hailuo line, following Hailuo 01 and 02, and the first with open weights. The company has also launched MiniMax M3, a new LLM, and MiniMax Speech 2.8 in recent weeks, signaling an acceleration across modalities.
Sources
