Since RuntimeWire's earlier H3 launch report, founder Yan Junjie's MiniMax has made two base checkpoints downloadable on Hugging Face and published a standard community license. The complete workflow behind MiniMax's 2K examples still requires hosted components.
Yan worked in computer vision before starting the company. The new files let developers inspect the local deployment requirements and clarify where MiniMax retains control of the production workflow.
What developers can download
The 288 GB Hugging Face repository contains two task-specific checkpoints. H3-Base-FL2VA handles text-to-video and generation from first, last, or first-and-last frames. H3-Base-Ref2VA accepts packages of reference images, video and audio.
Both checkpoints produce output with a 768-pixel shorter side. MiniMax's deployment example uses four GPUs through SGLang. The initial release supports full-attention inference, while the sparse-attention implementation remains unpublished.
What remains hosted
H3-Context-IR is not included in the download. MiniMax says the component depends on multiple models and services and contributes to final output quality.
H3-Regenerate-2K also remains outside the repository. According to the model card, it takes a 768p result and the original context, then regenerates the clip at higher resolution. Reproducing the workflow used for MiniMax's official 2K examples requires a local H3-Base deployment combined with MiniMax's hosted services through its API.
Where to check license terms
MiniMax publishes the MiniMax H3 Community License alongside the repository. Developers should review the license terms before using, modifying, distributing or displaying H3 weights or outputs.
The checkpoint release therefore gives developers who can comply with the standard license access to MiniMax's base video-generation models and local deployment instructions. It does not provide a fully local version of the 2K workflow, which continues to depend on MiniMax's API.