What H3 Actually Is

One name, three modules

MiniMax H3 is a unified multimodal video generation model. You send it text — optionally with images, video clips, or audio — and it returns finished video, with sound. On the API it goes by exactly one name: MiniMax-H3. That string is case-sensitive, and it is the only value the model field accepts on every endpoint you will use in this course.

Under that one name sit three logical modules, and keeping them separate in your head will pay off in almost every later lesson:

• H3-Context-IR — the prompt and context understanding stage. It takes what you wrote and expands it into the detailed, structured prompt the generator actually consumes.
• H3-Base — the generator itself: a 33-billion-parameter dense model MiniMax calls the H3-Omni-Transformer. This is what produces the video.
• H3-Regenerate-2K — the upscaling stage. Instead of bolting on a separate super-resolution model, it uses H3-Base itself to regenerate a 768P output at 2K, in context.

For architectural color: text and vision enter through an encoder built on Qwen3-VL-32B; visual latents pass through a VAE with 16× spatial and 4× temporal compression and 24 latent channels; audio has its own VAE working in 32 kHz stereo with dual independent channels. You will never touch these components directly through the API — but the fact that the model has an audio VAE at all sets up the next section.

The headline most guides missed: it makes sound

H3 generates synchronized stereo audio natively — 32 kHz stereo, produced as part of the video output in the same generation, not stitched on afterwards. That includes on-screen dialogue with lip-sync, with stable support in 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish. Support extends beyond those eleven, but less stably. Output video runs at 24 FPS.

This changes what the model is for. A talking character, a product spot with a voiceover, ambient sound that matches what is on screen — these are single-generation jobs, not video-plus-audio-pipeline jobs. The paid core of this course spends a full lesson on driving dialogue and sound; the capstone requires dialogue in a non-English language. For now, just recalibrate: H3 is an audio-video model.

Open source — read the module list, then the license

On 2026-08-03, MiniMax published the announcement "Open General Intelligence: MiniMax H3 Is Now Open Source" and put weights on Hugging Face at `MiniMaxAI/MiniMax-H3`. That much is true and current — this course exists partly because that release made every earlier guide stale overnight.

But "H3 is open source" needs two qualifications before you repeat it.

Second, the license. The weights ship under the MiniMax H3 Community License Agreement — a custom license, not an OSI-approved open-source license, and it does not grant rights everywhere. The LICENSE file defines an "Applicable Territory" of "worldwide, excluding the Excluded Territories," and names those Excluded Territories as the European Union, the United Kingdom, the Republic of Korea and the United States of America.

So when your legal team asks "is it open source?", the accurate answer is four things at once: open weights, custom license, partial pipeline, territory-restricted. The paid core has a full lesson (h3-c9) on what self-hosting H3-Base realistically involves, the territory gate in detail, and when it beats paying per second.

Three ways to ask for a video

The API offers three generation modes. The official docs use short-form names you will see in code comments and in this course:

• Text-to-Video (`t2va`) — a prompt alone. The model invents everything.
• First/Last-Frame Image-to-Video (`i2va`) — you supply the exact first frame, the exact last frame, or both, and the model generates the motion. With both frames set, you get a controlled transition between two images you chose.
• Reference Generation (`r2va`) — you supply reference images, reference video clips, or reference audio, and the model keeps the identity of what you showed it — a character, a style, a motion, a voice — without being pinned to exact frames.

A prompt is required in every mode. How the modes differ in their rules — and the important fact that frame control and reference control cannot be mixed in one request — is lesson h3-f4 and the core track.

Why "verified against the docs" is the whole pitch

While building this course we checked every claim in the most widely circulated H3 write-ups against the official MiniMax documentation, API schemas, and announcement. The audit found 25 material errors — wrong endpoint paths, a request parameter that does not exist, a task status string the API never returns, missing input constraints that cause real failures, and the entire native-audio capability missing. The worst of them would break your code; the subtlest would quietly corrupt your mental model.

Those corrections are woven through every lesson as callouts like the two above, and the capstone certification checklist requires you to clear all of them. When this course states a number, an endpoint, or a parameter name, it traces to an official MiniMax source. When something is not documented, we say so instead of guessing.

Where you go from here

The rest of the free foundation gets you productive: h3-f2 covers accounts, keys, and exactly what generation costs; h3-f3 walks you through your first real API call end to end; h3-f4 gives you the complete constraint table so your requests stop failing before they start. The paid core then goes deep on the content-array role system, camera and prompt grammar, native audio, Context-IR, the 2K regeneration pipeline, production operations, and self-hosting the open weights.