MiniMax H3

MiniMax's 33B open video model makes video and stereo sound in one pass, up to 15 s, but its license excludes the US, EU, UK and Korea.

In short. MiniMax H3 is an AI video model whose weights you can download. From text, pictures, video clips or sound it makes a clip of up to 15 seconds with matching stereo audio, including speech. It is very large, and its license does not cover the United States, the European Union, the United Kingdom or South Korea, so check where you and your users are before you rely on it.

Summary

MiniMax released H3 as open weights on 3 August 2026; the license is dated 2 August. It is an omni-modal generator: text, images, video and audio can all be used as context, and the model produces video together with native 32 kHz stereo audio in one pass. Clips run 4 to 15 seconds at 24 frames per second, with a default short side of 768 pixels, and MiniMax lists stable dialogue in 11 languages. Two BF16 checkpoints are published. FL2VA covers text-to-video and first-frame, last-frame or both-frame conditioning. Ref2VA covers reference-based generation from up to 9 images, 3 video clips and 3 audio clips, with at most 12 files in total. Each checkpoint is a separate service, and both have classifier-free guidance distilled in.

The core is the H3-Omni-Transformer, a 33-billion-parameter dense, single-stream transformer that predicts video and audio latents together. About 13 billion of those parameters sit in AdaLN branches whose outputs can be precomputed, so MiniMax says they do not need to be loaded for inference-only use; the full weights are released to allow fine-tuning. Text and visual context pass through the Qwen3-VL-32B text encoder, which is Apache 2.0, and separate video and audio autoencoders handle pixels and sound. The release includes full-attention inference only; MiniMax says its sparse-attention version will follow.

Two parts of MiniMax's complete system are not in the download: H3-Context-IR, a hosted step that rewrites complex multimodal prompts, and H3-Regenerate-2K, which regenerates the 768p result at 2K. MiniMax offers both through its API, so local use means 768p output. The Hugging Face repository is about 498 GB because it keeps the original and diffusers-format files side by side; one checkpoint with its text encoder and autoencoders is about 144 GB. MiniMax's self-hosting guide uses an 8 x B200 reference setup and publishes no minimum memory, while the SGLang benchmarks it cites show a 5-second 1344 x 768 clip in about 75 seconds on 4 x H200 GPUs. SGLang also lists offloaded recipes for two RTX 5090 cards and one 24 GB RTX 4090, which MiniMax does not verify. Besides SGLang, the model card points to diffusers, vLLM and ComfyUI for running it.

Good for

  • Short clips with synchronized speech and sound generated locally, without per-clip fees
  • Reference-driven work that mixes several images, video clips and audio samples in one prompt
  • Studios and research teams with data-center GPUs that want to fine-tune an audio-video model
  • Projects outside the European Union, the United Kingdom, South Korea and the United States

Watch out for

  • The license does not cover the European Union, the United Kingdom, South Korea or the United States, so people there need a separate license from MiniMax
  • Commercial products earning more than US$20 million a year need MiniMax's written authorization, and any commercial product must show "MiniMax H3" in its interface
  • Local output is 768p; the 2K step and the prompt-rewriting step are available only through MiniMax's API
  • One checkpoint is about 144 GB, and MiniMax publishes no minimum GPU memory, so plan for several GPUs or accept slow offloaded runs
  • The license forbids using the model or its outputs to train other AI models

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of MiniMax H3. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
Base FL2VA (BF16) BF16 33.1B 144 GB 166 GB Big-memory desktop
Base Ref2VA (BF16) BF16 33.1B 144 GB 166 GB Big-memory desktop

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

A community license that lets you use, modify and distribute MiniMax H3, but only in the "Applicable Territory": the whole world except the European Union, the United Kingdom, South Korea and the United States. Commercial products earning over US$20 million a year need MiniMax's written authorization, every commercial product must show "MiniMax H3" in its interface, and you may not use the model or its outputs to train other AI models. Get legal advice, especially if you or your users are in an excluded region.

  • Attribution or model name must be shown
  • Free commercial use only below a revenue cap
  • Not licensed in some regions
  • Acceptable-use policy applies
  • Outputs must be filtered or reviewed, and AI generation disclosed

More about MiniMax H3 Community

See also

Frequently asked questions

What is MiniMax H3?
MiniMax H3 is an AI video model whose weights you can download. From text, pictures, video clips or sound it makes a clip of up to 15 seconds with matching stereo audio, including speech. It is very large, and its license does not cover the United States, the European Union, the United Kingdom or South Korea, so check where you and your users are before you rely on it.
Can I use MiniMax H3 commercially?
Yes, with conditions. MiniMax H3 is released under the MiniMax H3 Community, which allows commercial use but adds extra rules: Attribution or model name must be shown; Free commercial use only below a revenue cap; Not licensed in some regions; Acceptable-use policy applies; Outputs must be filtered or reviewed, and AI generation disclosed. Read the license text before you ship a product.
Is MiniMax H3 open source?
MiniMax H3 is open-weight, not open source in the strict sense. You can download the model files, but the MiniMax H3 Community adds conditions that the Open Source Initiative definition does not allow.
Can I run MiniMax H3 on my own computer?
Only on a machine with a lot of memory. The smallest variant, Base FL2VA (BF16), needs about 166 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
When was MiniMax H3 released?
MiniMax H3 was released by MiniMax on August 3, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
LTX-2.5 LTX (Lightricks) 22B – LTX-2.x Community (Conditional) 16 GB Laptop or phone
Wan 2.2 Alibaba Wan 5B – 27B 14B active – Apache 2.0 (Permissive) 24 GB Gaming GPU
MiniMax H3-Image MiniMax 33.1B – MiniMax H3-Image Community (Conditional) 155 GB Big-memory desktop
Pika H3 Flex Pika 33.1B – MiniMax H3 Community (Conditional) 169 GB Big-memory desktop