The useful answer

  • H3 accepts mixed text, image, video and audio context.
  • The base model can produce video with native stereo audio.
  • Local 768p and hosted high-resolution workflows are not identical.
  • Cost comparisons should include retries, duration and resolution.

What MiniMax H3 is

MiniMax H3 is a general-purpose, omni-modal generation system. Instead of treating text-to-video, image-to-video and reference-video generation as isolated products, it reads a mixed context and generates synchronized audio-visual output.

The official materials describe a 33B-parameter H3-Omni-Transformer, separate visual and audio latent systems, 24 fps output and native 32 kHz stereo audio. A request can produce between four and fifteen seconds of output, depending on the interface and workflow used.

Open weights versus the full hosted workflow

The open release centers on H3-Base and includes task-specific checkpoints for first/last-frame and reference-driven generation. The official 2K result also uses hosted context interpretation and an in-context regeneration stage. Those hosted pieces are not the same thing as downloading the base checkpoint.

That distinction matters when comparing demos. A local 768p result, a hosted H3 result and an H3 Max endpoint can differ in resolution, preprocessing, speed and price even when they share the H3 name.

  • Use hosted generation when setup time and predictable per-clip billing matter most.
  • Use open weights when control, repeatable workflows or private infrastructure justify the hardware work.
  • Document the exact checkpoint, workflow and provider before comparing output quality.

The production-cost checklist

A quoted per-second rate is only the base layer. A useful H3 budget multiplies duration by the provider rate, then accounts for resolution, variants, discarded generations and any reference-input charges.

For a content pipeline, separate experimentation from approved output. Prompt tests may use short clips or a lower resolution, while approved shots receive the expensive pass. The calculator on this site models both the direct clip cost and retry overhead.

Where to go next

Choose the guide that matches your intent. ComfyUI users should start with the installation and workflow articles. Developers evaluating downloadable artifacts should use the Hugging Face, GitHub and GGUF guides. Teams choosing between hosted endpoints should read the H3 Max comparison and run both scenarios through the calculator.

Questions answered

What people ask about minimax h3

Is MiniMax H3 open source?

MiniMax publishes model weights and inference code under its H3 Community License. Some hosted context-processing and high-resolution workflow components are not part of the open release.

Does MiniMax H3 generate audio?

Yes. H3 jointly generates video and native stereo audio rather than requiring a separate sound-generation pass.

Can MiniMax H3 run locally?

Yes, H3-Base can run locally through supported runtimes, but the full hosted 2K workflow includes components beyond the downloadable base model.

Primary reading

Sources behind this guide

MiniMax H3 model overviewFirst-party product overview and supported modalities.MiniMax H3 official model cardWeights, architecture, capabilities and license links.MiniMax H3 official repositoryReference implementation, checkpoints and prompt guidance.