MiniMax H3 Brings Open Weights to 2K AI Video, Audio Included
All blog articles
A 2K video generator with native stereo sound, free to download: that is the rock MiniMax just threw into the pond. The Shanghai lab has officially opened the weights of H3, days after its July 31 launch. Veo 3.1, Kling 3.0, and Sora 2 suddenly have a free rival.
MiniMax H3 hears the scene before it films it
H3 reads text, images, video, and audio as one shared context. The model then generates clips of 4 to 15 seconds, with dialogue, sound effects, and room tone included, in a single pass. Everything runs at 24 frames per second, with stereo sound in 11 languages.
Audio even works as an input: a soundtrack can drive the visual rhythm of a scene. For control, H3 accepts up to nine reference images, three videos, and three audio tracks. A written instruction can also edit an existing clip.
Open weights, but not the whole system
One useful nuance: the module published on Hugging Face is H3-Base, a 768p generator with audio. The Context-IR orchestration layer and the full 2K regeneration pipeline stay hosted on MiniMax's servers.
The real weapon is price. At 2K, MiniMax claims a per-second cost under a third of mainstream models. At 768p, the bill reportedly drops below half of comparable offers.
MiniMax H3 against Seedance, Veo, and Kling
The timing is no accident: ByteDance shipped Seedance 2.5 within the same 24 hours. Chinese labs now hunt in packs, on quality and on price alike. A MiniMax engineer even claims the top spot for video editing with audio on Artificial Analysis, a boast still to be verified.
After Kimi K3 on the text side, China's open-weight wave now reaches video. One caveat: the hosted service is not offered in the US, EU, or UK, due to regulatory and copyright questions. The weights themselves travel visa-free.
Video is open source's next playground
MiniMax also points at robotics: its researchers describe video generation as an engine for world models. Check back in a few weeks, when the first community forks will show whether open-source AI video holds up.
Is MiniMax H3 really free and open source?
The H3-Base generator weights download freely from Hugging Face and run locally, subject to the license terms. The Context-IR layer and the hosted 2K rendering still require MiniMax's paid API.
How long can a MiniMax H3 video be?
MiniMax H3 generates clips of 4 to 15 seconds at 24 frames per second, up to 2K, with synchronized native stereo audio. Aspect ratios include 16:9, 1:1, 3:4, and 9:16.