MiniMax founders Yan Junjie and Yeyi Yun are preparing to expand the downloadable parts of H3, the Shanghai-based AI developer's multimodal video model. In an August 7, 2026 community discussion, H3 team members said MiniMax plans to release its 2K regeneration stage, publish sparse-attention reference code and develop a unified model for image generation and editing.
MiniMax's official recap The H3 discussion
The roadmap concerns the gap between MiniMax's hosted developer platform and the components developers can download and inspect. MiniMax did not provide release dates or final license terms for the planned components.
MiniMax was founded in early 2022. Yan is founder, chairman, chief executive and chief technology officer, while Yun is co-founder and chief operating officer. According to MiniMax's management biographies, Yan spent more than six years at SenseTime and became a vice president and vice-head of its research institute. He studied mathematics at Southeast University, earned a doctorate in artificial intelligence from the Chinese Academy of Sciences and conducted postdoctoral research at Tsinghua University.
Yun previously worked at SenseTime in fundraising, strategic investment and corporate strategy. MiniMax says she now oversees operations, commercialization, product strategy and business development. Yan has described MiniMax's objective as making general-purpose AI affordable and useful for creative and productive work in a 2025 WAIC keynote.
MiniMax launched H3 on July 31, 2026. MiniMax describes H3 as accepting text, images, video and audio as context and generating video with native stereo audio, including dialogue, music and ambient sound. The launch post describes text-to-video, image-to-video, video-to-video motion transfer, reference-based generation and controllable editing. H3 clips can run for up to 15 seconds, and MiniMax advertises output at up to 2K resolution.
What developers can download now
MiniMax initially published H3 base checkpoints that produce 768p output. MiniMax's model documentation lists H3 at 768p and 2K, but the initial downloadable weights use full attention and do not include the separate regeneration stage used to produce the 2K output advertised for MiniMax's hosted H3 service.
The current release therefore gives developers the 768p base model for local inspection and testing. H3-Regenerate-2K and sparse-attention reference code remain planned releases, while the unified image model remains under development. MiniMax has not established when those assets will arrive or what licenses will govern them.
The 2K model adds a regeneration stage
MiniMax says it plans to release H3-Regenerate-2K, a separate stage intended to extend the publicly released 768p model to 2K output. The commitment appeared in the H3 team's community discussion.
Releasing that component would let developers examine the additional computation, memory requirements and output changes associated with H3's 2K process. Until then, developers cannot reproduce MiniMax's highest-resolution H3 workflow solely with the public base checkpoints.
The planned release would also expose a central part of H3's commercial pitch. MiniMax advertises output of up to 2K and says its 2K generation costs less than one-third as much per second as mainstream alternatives. The supplied research contains no independent benchmark or price table that reproduces MiniMax's comparison.
Sparse attention remains a future optimization
The open weights initially use full attention. MiniMax says a sparse-attention reference implementation is planned, but its performance and hardware requirements remain untested. The company described the planned code as a conservative reference implementation in the H3 community discussion.
A MiniMax-authored implementation would give developers a documented starting point for measuring memory use, generation speed and output quality against the full-attention checkpoints. Any gains will depend on the code MiniMax releases, its hardware support and the evaluation material used. The community statements establish a roadmap, while the performance implications remain to be tested by developers.
A related image model is planned
MiniMax plans to develop a unified model for text-to-image generation and general-purpose image editing, according to the community discussion. MiniMax has not provided a release date, license or reproducible evaluation for the model.
One H3 researcher reported strong zero-shot results from image-editing tests of a model trained to predict a final frame from an initial frame and a caption, without editing-specific post-training. That account remains an experimental result reported by the H3 team and has not been independently reproduced in the supplied material.
In a screenshot of the discussion, an account identified as ryan85127704 replied, "Yep, Keep open until AGI arrived." The supplied material does not establish that the account belongs to a MiniMax employee.
The reply expresses an intention; it does not set a release schedule or license commitment. Each planned release will be defined by the weights, code, documentation and legal terms MiniMax eventually publishes.
MiniMax brings public-market funding to H3
MiniMax raised approximately $619 million in a January 2026 Hong Kong initial public offering. MiniMax's listing prospectus named cornerstone investors including Alibaba-linked Alisoft, Aspex, IDG Breyer, Eastspring, E Fund, Janchor, Martis, Mirae Asset, Perseverance and Taikang.
Alibaba led a $600 million financing round in March 2024 that valued MiniMax at approximately $2.5 billion. The research brief also names HongShan, Tencent, IDG Capital and Hillhouse among MiniMax's backers.
MiniMax reported $79 million in 2025 revenue, an increase of 158.9% from 2024, with more than 70% generated outside China. Its full-year results announcement reported $252.8 million in research and development spending, a $250.9 million adjusted net loss and $1.0503 billion in cash as of December 31, 2025.
MiniMax says its products have cumulatively served more than 300 million individuals and more than one million enterprises and developers. Those totals cover MiniMax's language, video, audio and music products rather than H3 alone.
Hailuo AI averaged 5.65 million monthly active users during the first nine months of 2025, up from 2.17 million in 2024, according to MiniMax's prospectus. MiniMax places its video products in competition with OpenAI's Sora, Google's Veo and Kuaishou's Kling.
H3's claimed distinctions are its multimodal inputs, native audio generation, reference conditioning and downloadable weights. The planned releases would let engineers test a larger share of that stack locally. H3-Regenerate-2K would address the current resolution gap, while MiniMax's sparse-attention implementation could provide an inference path for independent measurement. Release dates, licenses and reproducible performance will determine how much of MiniMax's hosted workflow developers can actually run themselves.