MiniMax Open-Sources H3 to Challenge ByteDance’s Seedance 2.0
Chinese artificial intelligence company MiniMax has open-sourced H3, a 33-billion-parameter multimodal video model designed to generate moving images and stereo audio in sync. The release challenges proprietary offerings including ByteDance’s Seedance 2.0 and could broaden access to advanced video generation by allowing developers and creators to deploy the model locally instead of relying solely on cloud-based services.
As of Aug. 10, 2026, H3 ranked first on Artificial Analysis’ video-editing leaderboard and had overtaken Seedance 2.0 in text-to-video performance, according to the reported results. MiniMax said the model can run offline on a consumer graphics card with 12GB of video memory, while its generation cost is about 57% of Seedance 2.0’s, giving H3 a potential advantage in pricing and developer adoption.
All Coverage
1 original reportsThe Backstory
The history behind this eventMiniMax Unveils H3 Omni-Modal Model, Pledges Open Weights
Shanghai-based artificial intelligence startup MiniMax is expanding beyond single-purpose media generators with H3, also branded Hailuo 3.0. The model combines text, image, video and audio understanding and generation in one system, while targeting commercial video editing and cross-modal reference workflows. Its significance lies in MiniMax’s effort to challenge proprietary flagship products through broader capabilities, lower computing costs and a planned open-weight release.
MiniMax unveiled H3 on July 31, 2026, saying it can produce videos of up to 15 seconds at 2K resolution with native stereo audio and use mixed-media inputs as references. The company said 2K generation costs less than one-third as much as mainstream comparable models, though it did not disclose an absolute dollar or yuan price. MiniMax also said the model weights would be released within days of the announcement.
Subscribe to Mark Radar Weekly
Every Friday, the week's strongest signals in your inbox. Unsubscribe anytime.