MiniMax H3 Is Open Weights — Here Is What You Actually Get

MiniMax announced H3 as an API product on July 31, 2026 and published the weights to Hugging Face three days later. It is a 33B-parameter system with native audio generation, and it is served today by SGLang, vLLM, diffusers and ComfyUI.

"Open weights" is doing a lot of work in that sentence. H3 is a three-module system and only the middle module shipped. Here is what you can actually run on your own hardware, what still calls MiniMax's servers, and the licence clause that decides whether any of this is available to you at all.

What H3 is

H3 is an omni-modal generative system: it takes text, images, video and audio as a single multimodal context and generates video with native stereo audio — the video and audio latents are predicted jointly by one transformer, not stitched together afterwards.

Spec	Value
Duration	4–15 seconds
Frame rate	24 fps
Resolution	768p default; 2K via H3-Regenerate-2K
Audio	32 kHz stereo, generated with the video
Aspect ratios	21:9, 16:9, 4:3, 1:1, 3:4, 9:16 and others
Dialogue languages	11 with stable support, including English, Chinese, Japanese, Korean, Arabic and Spanish

Inputs come in two mutually exclusive shapes, which is the detail most summaries drop:

• H3-Base-FL2VA — frame conditioning. Zero, one or two images: none is text-to-video, one is a first or last frame, two is first-and-last.
• H3-Base-Ref2VA — omni-reference. Up to 9 images, 3 video clips and 3 audio clips; every clip must be 2–15 seconds, and no more than 12 files across all types.

You pick one. Frame conditioning and reference inputs cannot be combined in a single request — they are separate model variants, served on separate endpoints.

The three modules, and the one you got

A request to the H3 API passes through three stages:

1. H3-Context-IR reads your free-form inputs — prompt, images, clips, audio — works out how they relate to each other and to the video you want, and serialises that understanding into a structured Context Intermediate Representation.
2. H3-Base generates 768p video and audio from that IR. This is the open-weights part.
3. H3-Regenerate-2K feeds the 768p result back in alongside the original context and regenerates at 2K.

Modules 1 and 3 are not in the release. MiniMax is explicit about why module 1 is missing — it "relies on a multi-stage workflow and multiple hosted models and services" — and equally explicit about what that costs you: H3-Context-IR "is critical to the quality of the final output, so we strongly recommend incorporating it into your generation pipeline."

That recommendation is not decoration. The IR is not a rewritten prompt; it is a structured document with named subjects, a retention analysis saying which parts of each reference must survive, a shot-by-shot description with dialogue inline, and a soundscape section. One published example runs past 33,000 prompt tokens for a five-second clip. You can build your own from the two prompt-writing guides MiniMax ships, but you are then rebuilding the component they identify as the main determinant of output quality.

The same is true of resolution. The card's own "Full 2K Workflow" is a local SGLang deployment calling two MiniMax hosted APIs around it. Self-hosted and offline, H3 is a 768p model.

The licence excludes four markets — read this before you plan around it

This is the part that changes whether the rest of the article applies to you. From the MiniMax H3 Community License Agreement:

"Applicable Territory" means worldwide, excluding the Excluded Territories. … "Excluded Territories" means the European Union, the United Kingdom, the Republic of Korea and the United States of America.

The open-weights grant does not extend to those four. Three consequences worth being precise about:

• It covers outputs, not just weights. The agreement prohibits use, reproduction, modification, distribution or display of the works "or any of their Outputs or results" outside the Applicable Territory. A video generated lawfully elsewhere is still not licensed for use in an excluded territory.
• Over $20M in yearly revenue needs written authorisation — a separate, prior approval from MiniMax, requested by email.
• There is a route in. Organisations in the excluded territories can apply for a formal licence; MiniMax reviews the deployment and the controls around it. Their stated reasoning is regulatory rather than commercial — the EU AI Act, UK and Korean rules on AI-generated content, and ongoing US copyright litigation over generative video — and they describe the limit as "not yet, not never."

The API is global. MiniMax draws the line at distribution, not geography: when they operate the serving infrastructure they can enforce their own safeguards, so the hosted model is available everywhere the weights are not.

What it takes to run

H3-Omni-Transformer is a 33B dense, single-stream transformer. Two details make the practical number smaller than that:

• Roughly 13B of those parameters sit in AdaLN branches whose modulation outputs can be precomputed and cached. Inference-only deployments do not load them. The full weights ship anyway, so fine-tuning stays possible.
• H3 natively supports sparse attention, but the initial open-source release is inference with full attention only. The sparse implementation is promised in a later update — until it lands, you are paying full quadratic cost on long multimodal sequences.

MiniMax's reference SGLang configuration serves each variant across four GPUs with Ulysses sequence parallelism, and the two variants run as separate services on separate ports. SGLang, vLLM, diffusers and ComfyUI all have documented paths, and the repository ships both the original checkpoint and a diffusers conversion so you only download the format your framework needs.

So: self-host or API?

Self-host if you are outside the excluded territories, you have four GPUs per variant to dedicate, and either 768p is genuinely enough or you are prepared to build your own context-processing stage. Fine-tuning is the strongest reason of all — full weights, and the licence permits derivatives within the territory.

Use the API if you are in the EU, UK, Korea or the US and have not applied for a licence; if you need 2K; if you want the official IR quality without rebuilding it; or if your volume does not justify standing GPUs. It is also the only version of H3 that is genuinely one model rather than one model plus an integration project.

H3 on Imgo

H3 is in Imgo's video generator as MiniMax H3, via MiniMax's official v2 API — the full three-module path, so 2K and the official IR are both included, and the territory question does not arise. Two constraints our model registry enforces, both inherited from the model rather than added by us:

• First and last frames cannot be combined with reference images, videos or audio. That is the FL2VA / Ref2VA split above, surfaced as a validation error instead of a failed generation.
• Reference video duration, and images past the first five, add upstream input charges.

H3 has no seed, so it is not the model to reach for when you need a shot to render identically next quarter — Veo 3.1 is the only one of the three majors that gives you one. What H3 gives you instead is the widest input surface of any model we route to: nine reference images, three reference videos and three reference audio clips in a single request, with the soundtrack generated alongside the picture.

If you want to compare it against Seedance 2.5 and Veo 3.1 before committing a pipeline, the three are one dropdown apart and share a request shape and a credit balance.