MiniMax H3.
MiniMax H3 is an open-weight video model from MiniMax. Specs, recommended settings, hardware and licence terms, from the developer's own pages.
- Status
- Reference
- Task
- Text to video, Image to video
- Input
- Prompt, optional first and last frames, or reference images, video and audio
- Output
- 4 to 15 s clip with stereo audio, 768p
- Developer
- MiniMax
- Parameters
- 33B
- Licence
- MiniMax H3 Community License
- Released
- August 2026
What the model does.
Paraphrased from the official model card.
MiniMax H3 is an omni-modal video model from MiniMax that generates video and native stereo audio together. It takes text, images, video and audio as context and produces 4 to 15 second clips at 24 fps with 32 kHz stereo sound, with dialogue in 11 languages.
The open release is H3-Base, a 33B single-stream transformer, as two checkpoints: one for text-to-video and first/last-frame video, one for references (up to 9 images, 3 video clips and 3 audio clips). It generates at 768p. The prompt preprocessor and the 2K regeneration stage behind MiniMax's own output aren't open; they're available only through MiniMax's API. The licence excludes the US, the EU, the UK and South Korea.
Reference page.
We don't currently serve MiniMax H3. The specs, settings and licence terms on this page come from MiniMax's own pages.
Is MiniMax H3 licensed for commercial use?
Licence: MiniMax H3 Community License.
Conditional. The licence doesn't cover the United States, the European Union, the United Kingdom or South Korea; organisations there have to apply to MiniMax for a separate licence. Elsewhere commercial use is allowed, but businesses with more than US$20 million in yearly revenue from their products need MiniMax's prior written authorisation, and commercial products must show "MiniMax H3" in their interface.
- The licence applies only in the Applicable Territory: worldwide, excluding the European Union, the United Kingdom, South Korea and the United States. Section V.4: using, reproducing, modifying, distributing or displaying the works or their outputs outside that territory isn't authorised. MiniMax-H3, LICENSE
- Organisations in the excluded territories can apply to MiniMax for a licence; the card links an application form for the US, EU, UK and South Korea. MiniMax's licence Q&A says the limit means "not yet", not "not ever". MiniMax-H3, QA-about-License.md
- Section IV.1: if your commercial products and services generate more than US$20 million in yearly revenue, you need prior written authorisation from MiniMax (api@minimax.io). Section IV.2: commercial products must display "MiniMax H3" prominently in their user interface. MiniMax-H3, LICENSE
- Section V.2 and V.5: before giving anyone access, bind each user to terms at least as protective as Section V and Exhibit A, run safeguards against violating uses, keep a way to report violations, and act on reports. MiniMax-H3, LICENSE
- Section V.3: the works and their outputs can't be used to improve any other AI model. MiniMax-H3, LICENSE
- Section III: anyone who receives the works or uses your related products or services must be given a copy of the agreement, and distributions other than hosted services need a NOTICE file. A "Powered by MiniMax H3" notice, AI-generation labels on files and a public write-up are encouraged, not required. MiniMax-H3, LICENSE
- Exhibit A bars, among other things, placing machine-generated content in public without a clear disclosure, impersonation without consent, military use and high-risk automated decisions. MiniMax claims no rights over outputs. Governing law is Hong Kong. MiniMax-H3, LICENSE
This is our reading of the licence, not legal advice. Read the licence text before you rely on it.
MiniMax H3 hardware requirements.
Published by MiniMax for running the weights yourself. These are the developer's figures, not Windpaint's.
- Published VRAM
- Not stated
- GPUs named
- 4 GPUs in the SGLang example
- Applies to
- No VRAM figure on the card. BF16 checkpoints with a Qwen3-VL-32B text encoder.
- The card gives no VRAM figure. Its SGLang deployment example serves each checkpoint with
--num-gpus 4 --ulysses-degree 4. MiniMax-H3 model card - The released checkpoints are CFG-distilled BF16 weights. Each one bundles the Omni Transformer with its processor, tokenizer, text encoder, visual VAE and audio VAE. MiniMax-H3 model card
- The text encoder is the full Qwen3-VL-32B, and H3 uses the hidden states from its 50th layer. H3's own tokenizer files are required, because H3 adds special tokens such as
<d>. MiniMax-H3 model card - About 13B of the transformer's 33B parameters are in AdaLN branches. Their outputs can be precomputed and cached, so the card says those parameters don't need to be loaded for inference-only deployment. MiniMax-H3 model card
- Recommended frameworks: SGLang, vLLM, diffusers (
ModularPipeline.) and ComfyUI, with official ComfyUI text-to-video and reference-to-video workflow templates linked from the card. MiniMax-H3 model cardfrom_pretrained("MiniMaxAI/ MiniMax-H3")
MiniMax H3 native resolution.
What the developer's own code and examples use.
- Output: 4 to 15 seconds at 24 fps, with 32 kHz stereo audio. Aspect ratios include 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16. MiniMax-H3 model card
- H3-Base generates at 768p, with the shorter side at 768 pixels by default. 2K output comes from H3-Regenerate-2K, which isn't in the open release. MiniMax-H3 model card
Recommended settings for MiniMax H3.
Steps, guidance and defaults from the developer's card and code, for running the weights yourself.
- Two checkpoints. FL2VA handles text-to-video (
t2va) and first/last-frame video (fl2va): text plus an optional first frame, last frame or both. Ref2VA (ref2va) takes up to 9 images, up to 3 video clips and up to 3 audio clips, each clip 2 to 15 seconds and 15 seconds in total, with at most 12 files. MiniMax-H3 model card - Dialogue has stable support in 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish. MiniMax-H3 model card
- The card strongly recommends running inputs through H3-Context-IR, MiniMax's hosted preprocessor, or building your own preprocessing from its prompting guide. H3-Context-IR isn't in the open release; MiniMax offers it through its API. MiniMax-H3 model card
MiniMax H3 prompt guide.
From MiniMax's own prompt guide, prompt tooling and examples.
- The guide's final prompt has three fields:
integrated_multimodal_description(visuals, action, shots, speakers, dialogue and diegetic sound along the timeline),overall_soundscape(ambient and action sounds, 1 to 4 sentences) andnon_diegetic_music(music only the audience hears, 1 to 3 sentences, or N/A). MiniMax-H3, VIDEO_PROMPT_WRITING_GUIDE_base_en.md - Image-conditioned prompts start with a fixed alignment line. For first-frame video: "For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced." MiniMax-H3, VIDEO_PROMPT_WRITING_GUIDE_base_en.md
- Number shots
[Shot 1],[Shot 2]and so on. The first has no timestamp; later shots begin with a cut time, such as "[Shot 2] At 00:03.500, the camera cuts to...". MiniMax-H3, VIDEO_PROMPT_WRITING_GUIDE_base_en.md - Write camera moves as motion type plus optional amplitude and speed, as a natural sentence: "The camera pushes in with small amplitude at slow speed toward...". Types include push in, pull out, pan, truck, tilt, pedestal, arc, tracking, static, shake, POV and roll. MiniMax-H3, VIDEO_PROMPT_WRITING_GUIDE_base_en.md
- Give each speaker a stable ID such as (S1) and put the spoken words inside
<d>with a language tag:<d>[English] I get off at the next station.</. Put on-screen text in English double quotes, word for word. MiniMax-H3, VIDEO_PROMPT_WRITING_GUIDE_base_en.mdd> - For first-and-last-frame video, the guide favours a single shot that describes the path between the two frames rather than the two frames themselves. MiniMax-H3, VIDEO_PROMPT_WRITING_GUIDE_base_en.md
MiniMax H3 limitations.
As the developer states them.
- H3-Context-IR and H3-Regenerate-2K aren't released, so a self-hosted H3-Base produces 768p output. MiniMax says it will release the 2K module once it's ready; the card's full 2K workflow uses MiniMax's API for those two stages. MiniMax-H3 model card
- The initial release runs full attention only. MiniMax says its sparse-attention implementation will follow in a later update. MiniMax-H3 model card
Questions.
Can I call MiniMax H3 on Windpaint?
Not at the moment.
Can I use MiniMax H3 in the US, EU or UK?
Not under its community licence. The licence excludes the United States, the European Union, the United Kingdom and South Korea, and use there isn't authorised. Organisations in those regions can apply to MiniMax for a separate licence.
Is MiniMax H3 open source?
No, it's open-weight. The H3-Base weights are public on Hugging Face under the MiniMax H3 Community License Agreement, not an open-source licence. The H3-Context-IR preprocessor and the H3-Regenerate-2K stage aren't included; MiniMax offers them through its API.
Can I use MiniMax H3 commercially?
Inside its licensed territory, yes, with conditions. Products earning more than US$20 million a year need MiniMax's prior written authorisation, commercial products must show "MiniMax H3" in their interface, and users must be bound to the licence's use restrictions.
How much VRAM does MiniMax H3 need?
The card doesn't give a VRAM figure. Its SGLang example serves each checkpoint across 4 GPUs, and the text encoder is the full 32B Qwen3-VL.
Does MiniMax H3 work in ComfyUI?
Yes. The card links a ComfyUI tutorial and official workflow templates for text-to-video and reference-to-video.
Can a self-hosted MiniMax H3 make 2K video?
Not on its own. H3-Base generates at 768p, and the 2K stage, H3-Regenerate-2K, isn't released. The card's full 2K workflow pairs a local H3-Base with MiniMax's API for that stage.
How do I prompt MiniMax H3?
MiniMax's guide uses three fields: a timeline description with numbered shots, camera moves, speakers and quoted dialogue; an overall soundscape; and background music. Image-conditioned prompts start with a fixed line saying which frame each picture aligns with.