MiniMax H3 Open Weights: 768p Only, US/EU Excluded

MiniMax H3 makes video and stereo audio in one model

MiniMax calls H3 an omni-modal system: it reads text, images, video, and audio together as a single multimodal context. What comes out is a clip of 4 to 15 seconds with a stereo soundtrack, and you steer it through one of two released checkpoints. A checkpoint is a saved set of trained weights, and each of these two does one job.

One 33B transformer predicts video and audio together

The generator at the center, H3-Omni-Transformer, is a 33B-parameter dense transformer that predicts the picture and the sound in the same pass. It works on latents, compact numerical stand-ins for frames and audio, and only at the end are those decoded into a video track and a stereo track. The model card describes no later stage where a separate sound model has to be lined up with the footage.

H3-Encoder, which reads the prompt, is Qwen3-VL-32B with its full pretrained weights. The transformer takes the output of that model's 50th layer.

About 13B of the 33B parameters belong to AdaLN (adaptive layer normalization) branches, which rescale the network's internal values according to the conditions it is given. MiniMax says their outputs can be computed ahead of time and cached, so a deployment that only generates clips does not have to load them. So the 33B headline overstates what such a setup loads.

Clips run 4 to 15 seconds at 24fps with 32 kHz stereo

Resolution is the figure to read carefully: the shorter side of the frame is 768 pixels by default, and the 2K in MiniMax's system overview depends on an extra step, H3-Regenerate-2K. That step is not part of the open weights.

The listed aspect ratios run from 21:9 widescreen down to 9:16 vertical, with 4:3 and square formats in between. For spoken dialogue, MiniMax claims stable support in 11 languages and says others work to varying degrees.

Two checkpoints, one accepting up to 12 reference files

MiniMax split the release into two checkpoints, and the one you pick decides how you steer a clip: by pinning its first or last frame, or by handing over as many as 12 reference files. Both output video and audio together.

FL2VA covers the familiar modes. With no image it works as text-to-video, one image fixes the opening or closing frame, and two images fix both ends while the model fills in the motion between them.

Ref2VA is the one to reach for when a clip must feature a specific character or product, since it takes images, video, and audio as references in a single request. When your reference images come in mixed pixel dimensions, resize images to one size before upload so each file is framed the same way.

The MiniMax H3 open weights stop at 768p

The MiniMax H3 open weights contain one of the system's three modules, so a setup that runs entirely on your own hardware tops out at 768p, the resolution MiniMax gives for H3-Base. The other two modules prepare the input and raise the result to 2K, and the company keeps both on its own servers.

Only the H3-Base module was released

You download H3-Base from the MiniMaxAI/MiniMax-H3 repository on Hugging Face, where the weights come in BF16, a compact 16-bit number format. MiniMax says it publishes the complete weights so that others can keep developing the model, fine-tuning included.

One optimization also stayed behind. H3's final training stage used native sparse attention, which lowers the cost of long sequences by attending to only part of the input, but the first open release runs full attention only. The model card says the sparse implementation will follow in a later update.

The 2K step and input preparation stay on the MiniMax API

H3-Context-IR, which prepares the input, and H3-Regenerate-2K, which lifts the output to 2K, are both missing from the open release, and MiniMax offers them through its API instead. Context-IR sits in front of H3-Base and rewrites a free-form mix of text and media into the structured input the generator expects.

MiniMax left Context-IR out because the module depends on a multi-stage workflow and several hosted models and services. The company still calls it critical to final output quality, and strongly recommends either building it into your pipeline through the API or writing your own preprocessing by following its Prompting Guidance.

Regenerate-2K works differently from an upscaler: instead of enlarging pixels, it hands H3 its own 768p result plus the original context and has the model generate the clip again at 2K. MiniMax says this recovers small text and fine detail that an upscaler would have to guess. As of September 29, 2026, the model card still lists the module as not yet open-sourced, with a release promised once it is ready.

The missing 2K step matters when you judge your own output. MiniMax's full 2K workflow combines a locally run H3-Base with its Open Platform API, and its reference clips come in 2K and 768p versions made through that API. Compare a fully local run with the 768p references, not the 2K ones.

If you only want to try H3 without hosting anything, the model card lists two online routes: MiniMax's API platform and its apps, which include the Hailuo AI web app and a desktop app. The license Q&A describes the API as available globally, with safeguards built in.

Local runs use SGLang, vLLM, diffusers, or ComfyUI

MiniMax recommends four frameworks for running H3-Base yourself: SGLang, vLLM, diffusers, and ComfyUI. SGLang and vLLM are serving engines that run the model as a server other programs can call, and the model card links a cookbook for the first and a recipe page for the second. Hugging Face's diffusers is a Python library for generative models, while ComfyUI lets you build the pipeline by wiring nodes together on screen.

ComfyUI also has an official Comfy tutorial and two workflow templates, one for reference-to-video and one for text-to-video. If you would rather not write code, start there. The repository keeps the original checkpoints in FL2VA and Ref2VA folders next to a diffusers-format copy, and you can download only what your framework uses.

Running H3 on your own machine still counts as use under MiniMax's license. Where that machine sits, and where the clips are shown, decide whether the license covers you at all.

The MiniMax H3 license excludes the US, EU, UK, and Korea

The MiniMax H3 Community License, dated August 2, 2026, applies worldwide except in four excluded territories: the EU, the UK, South Korea, and the US. The agreement states at the top that its scope is limited to the Applicable Territory, and the licensor it names is Nanonoble.

Japan sits inside the applicable territory

Japan is not on the excluded list, so use in Japan falls within the Applicable Territory. The territory is defined as the whole world minus those four, so every other country is inside it as well.

The clause that matters in practice concerns where the model and its outputs are used. Section V.4 bars running, copying, modifying, or distributing the model outside the territory, and the same bar covers displaying what the model generates. Read literally, that means a team in Japan showing generated clips to viewers in the US or the EU would be acting outside what the license allows.

MiniMax says the exclusion means "not yet"

In its license Q&A, MiniMax says the current limit means "not yet", not "not ever", so it presents the exclusion as a matter of timing. The reason it gives is regulation. Video models, it argues, face a more complex and faster-changing set of rules than text or code models.

The Q&A then goes region by region. It notes that enforcement of the EU AI Act has started while the practical requirements for video and likeness generation are still taking shape. The UK and South Korea, it says, carry similar uncertainty, and in the US MiniMax is itself involved in ongoing copyright proceedings over generative video AI.

MiniMax says it had two options, holding the weights until every jurisdiction settled its rules or releasing now under a clearly scoped license, and it chose the second. Organizations in the four excluded territories can apply for a formal license, and MiniMax may approve use after reviewing the deployment and its safeguards. The API, it adds, is available globally with built-in safeguards.

Commercial use needs visible credit and approval above $20M

Commercial use is allowed inside the territory, but the product must credit "MiniMax H3" prominently in its user interface. Businesses whose commercial products and services pass $20 million in yearly revenue also need the licensor's written approval first, requested at api@minimax.io under a subject line the license spells out. Running H3 in the backend of a paid service with no visible credit would not meet the display clause.

Put together, the H3 release is a set of free weights for a 768p audio-and-video generator under a license that excludes the US, EU, UK, and South Korea, with the 2K and input-preparation steps kept on the API. Before building on it, check where your users and viewers are, whether 768p is enough without the API steps, and whether your yearly revenue tops $20 million.

Open-weight text models ship under different terms. Compare GLM 5.2 benchmarks, pricing, and Claude Code setup with DeepSeek V4 Flash 0731: Benchmarks and MIT License.