One model name, three production stages
The MiniMax H3 model card describes the audio-and-video generation trunk as a dense transformer with 33 billion parameters. About 13 billion parameters sit in AdaLN branches; because their outputs can be precomputed, they do not need to be loaded for inference-only deployment. The same page specifies output lasting 4 to 15 seconds, 24 frames per second, 32 kHz stereo audio, and a default image whose short side is 768 pixels. Together, those figures define concrete boundaries for a system an outside user can deploy.[1]
The crucial methodological distinction lies in the three-part production chain. H3-Context-IR converts multimodal input into an intermediate representation for the base model; H3-Base produces 768-pixel audio-video output; and H3-Regenerate-2K processes that result again with the original context to raise the resolution. The card provides three reproducible cases and scripts for the local base model. The end-to-end 2K workflow, however, combines local H3-Base with MiniMax's online interface. A locally reproducible result and the complete service-dependent chain are therefore different experiments.[1]
What a reference output cannot measure
The card's strength is that it exposes the stages through which an output travels. Its weakness is that the broad language about generalisation and quality is not tied to a comparative performance table. The scripts help reproduce a specified example, but they do not compare models under the same prompt, attempt count, hardware budget, and preregistered scoring rule. A difference in the result could come from the base transformer, the context intermediate representation, prompt construction, or the 2K regeneration stage. Assigning the gain to one component would be premature until those competing explanations are separated.[1]
The release supplies an important rung on the evidence ladder. Weights, runnable scripts, and explicit names for the three stages make externally testable questions possible. An informative study publishes fixed multimodal prompts and repeat counts in advance, then scores the local 768-pixel base output, online context processing, and 2K regeneration separately. Including failed runs reveals the distribution instead of only a selected best example. For H3, the firmest conclusion available today is which component can be reproduced where.[1]