MiniMax launches H3 2K video model with promised open weights
MiniMax launched H3 with 2K video and native stereo audio. MiniMax says the weights are coming in days, but the H3 license and official local hardware requirements remained unpublished as of August 2.
By Ryan Merket · Published
Why it matters
Usable H3 weights could let developers self-host, fine-tune and optimize a native audio-video model instead of relying entirely on MiniMax's API. The license, checkpoints and hardware requirements will determine how accessible that option becomes.
Yan Junjie's MiniMax, the Shanghai model maker, launched H3 on July 31 as a multimodal video model capable of generating clips with stereo audio at up to 2K resolution. MiniMax's H3 launch post says MiniMax plans to open the model's weights "in the coming days," subject to applicable laws and regulations.
The scope of that planned release remains unsettled. No completed public H3 weights release was verified as of August 2. MiniMax had not published an H3 license or official local hardware requirements in its launch post or API documentation. Those terms will determine whether developers can modify and deploy H3 on their own infrastructure and what commercial uses MiniMax permits.
Yan, MiniMax's founder, chairman, CEO and CTO, built his career in computer-vision research before starting MiniMax. He spent more than six years at SenseTime, including as a vice president and vice-head of its research institute. Yan studied mathematics at Southeast University, earned an AI doctorate from the Institute of Automation at the Chinese Academy of Sciences and completed postdoctoral research at Tsinghua University. MiniMax's management page credits him with about 200 academic papers and more than 30,000 citations.
H3 extends Yan's technical agenda beyond a conventional text-to-video or image-to-video generator. MiniMax designed the model to accept text, images, video and audio as context, then generate video with native stereo sound. MiniMax says H3 can produce clips of up to 15 seconds at 2K resolution. MiniMax's developer documentation lists 768p and 2K output, durations from four to 15 seconds and 24 frames per second.
A 2K launch with an open-weight test ahead
MiniMax describes H3 as an open, general-purpose multimodal video model. For now, the verified release is MiniMax's hosted model listing, not downloadable checkpoints. Downloadable files, license terms and official hardware guidance will determine the practical breadth of the promised open-weight release.
The model's central technical feature is its handling of several input types within one video system. MiniMax says H3 can jointly understand context spanning text, images, video and audio. Its developer documentation lists text-to-video, image-to-video, first-and-last-frame generation and multimodal reference generation among the supported modes.
MiniMax has released open model files before. MiniMax-M3 is downloadable from Hugging Face, giving developers one precedent to compare against H3 once its files and terms are available.
Yan's distribution bet
Usable H3 weights could carry the model into creative software, internal production systems and specialized computing environments beyond MiniMax's hosted API. Developers could test deployment costs, optimize inference and build fine-tunes if the license and released files permit those uses.
Lightricks provides one practical comparison. The official LTX-Video repository is already public on GitHub. H3's distribution value will depend on whether MiniMax supplies a similarly workable path from checkpoint to deployment.
The H3 launch follows a rapid sequence of financing and product announcements. Investing.com reported in July 2026 that MiniMax had closed roughly $2 billion in financing. Yan then teased H3 as a unified multimodal system on July 30, ahead of the July 31 launch.
MiniMax has the revenue base to pursue hosted sales alongside broader model distribution. MiniMax's 2025 annual report recorded $79 million in revenue, up from $30.5 million in 2024. The research brief says 73% of total revenue came from outside mainland China. International developer adoption therefore has direct commercial relevance for MiniMax.
MiniMax also says H3 costs less than one-third as much per second at 2K as what it calls mainstream models. Its launch post does not identify the comparison set or publish a reproducible pricing table. An open-weight release would shift part of the cost analysis to users' hardware, making memory requirements, inference speed and deployment tooling central to H3's economics.
MiniMax has set a timeline of days for opening H3's weights. The checkpoint, license and inference materials will establish whether outside teams can adapt the 2K audio-video model in practice and how H3 compares with open projects that already provide a documented local deployment path.