What is MiniMax H3? Multimodal video generation and its relationship to Hailuo
MiniMax H3 combines audiovisual references and text instructions to generate video with native stereo audio. Learn how it differs from older Hailuo models, how base weights work with services for 2K video and which community license applies.
Meet this model
MiniMax H3 is a multimodal generation system that interprets text and audiovisual references and produces video with native stereo sound. It follows the development of Hailuo 01 and 02 but is not simply a new alias for them. Hailuo AI is a creative product interface; H3 is a model system accessible through relevant services or open weights.
Use language to explain relationships among references
H3 is designed to handle relationships among modalities within one generation task. Given a product image and a motion clip, specify that appearance comes from the image and movement from the video. Explain each input’s role instead of merging unrelated subjects into one ambiguous description.
This suits adapting existing creative material but does not guarantee exact appearance and movement together. Compare product parts before and after occlusion and listen for sound-action alignment. No sample was generated here, and official demonstrations are not presented as local results.
Endpoint-frame and multi-reference branches take different inputs
The open release describes FL2VA and Ref2VA base branches. FL2VA accepts zero to two images for text-only or endpoint-frame conditioning; Ref2VA handles multimodal references. Select weights or APIs for the input mode rather than interchange their fields.
The release specifies video from 4 to 15 seconds with stereo audio. Fit dialogue length to the shot rather than squeeze a long line into a short clip. Playback is needed to establish intelligibility; successful submission does not establish clear speech.
The open base model differs from the complete 2K workflow
The open release lists a default 768-pixel short edge for base output and H3-Regenerate-2K for 2K regeneration. Reproducing the full workflow also involves contextual processing. Downloading base weights does not establish equivalence with the complete hosted 2K pipeline.
Before local deployment, list the local inference steps separately from dependencies on official services, then compare the final file. This establishes hardware needs and external dependencies. This article checked the release and repository without running inference or verifying a fully offline workflow.
Open weights remain subject to the H3 community license
H3 uses the MiniMax H3 Community License Agreement. Open weights do not imply unconditional permission. Read the license and acceptable-use requirements for model redistribution or hosted services and check the obligations for the intended usage.
Record older Hailuo model IDs separately from H3 branches. During migration, use the same authorized material to check accepted inputs and inspect video and audio. A successful old-model submission is not evidence that the new model behaves identically.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.