What Happened
On August 3, 2026, MiniMax announced the public weight release of H3-Base, following the broader H3 product announcement. Two checkpoints support text and first/last-frame generation, or generation from multimodal references. The release uses the MiniMax H3 Community License.
Why It Matters
H3 broadened access to models that jointly generate video and sound from richer context than a text prompt alone. The available checkpoints allow local deployment and adaptation, while the release also illustrates why an open-weight component should not be mistaken for an entirely open production system.
Technical Details
A 33-billion-parameter dense Omni-Transformer jointly predicts video and audio latents. Local H3-Base output is 768p. The complete service can produce up to 15 seconds with stereo audio and 2K output, but the Context-IR preprocessing system and Regenerate-2K module were not included in this release. The initial downloadable inference implementation uses full attention.