MiniMax H3 is MiniMax's video model: it makes 4 to 15 second clips with stereo sound from text, images, video and audio. Its open H3-Base weights run in ComfyUI 0.30.0 or later from built-in templates that download about 42 GB and render at a 768-pixel short side; 2K needs MiniMax's hosted stage. The MiniMax H3 Community License excludes the United States, the European Union, the United Kingdom and South Korea, so creators there use the Hailuo app, the MiniMax API or Comfy Cloud. The API lists H3 at $0.08 a second at 768P and $0.13 at 2K.
MiniMax H3 is a video model from MiniMax that turns text, images, video and audio into a clip of 4 to 15 seconds with stereo sound, and its open H3-Base weights run in ComfyUI at a 768-pixel short side. The license on those weights excludes the United States, the European Union, the United Kingdom and South Korea, where the hosted routes are the Hailuo app, the MiniMax API and Comfy Cloud.
Read on 13 October 2026 from MiniMax's model card, license, launch post and API pricing, Comfy's H3 guide, day-0 post and repackaged files, and the Hailuo plans page. Prices are in US dollars.
Quick answer
- Open weights: H3-Base, on Hugging Face since 3 August 2026. It renders at a 768-pixel short side. The 2K stage is hosted only.
- License: not valid in the US, EU, UK or South Korea. Elsewhere commercial use is allowed, with conditions.
- ComfyUI: native since version 0.30.0. Built-in templates, about 42 GB of files.
- VRAM: no published minimum. Comfy says offloading lets it run on a GPU like the RTX 3060.
- Hosted prices: $0.08 a second at 768P and $0.13 at 2K on the API. Hailuo plans start at $10.50 a month.
What is MiniMax H3?
MiniMax launched H3 on 31 July 2026 as a "general-purpose multimodal generation model" and published the weights on 3 August. Comfy's day-0 post calls it the first video model the company has released with open weights. The model card gives these specifications:
| Spec | MiniMax H3 |
|---|---|
| Clip length | 4 to 15 seconds |
| Video and audio | 24 frames a second with 32 kHz stereo audio, generated together |
| Resolution | 768-pixel short side by default. 2K comes from the H3-Regenerate-2K stage |
| Dialogue | Stable support for 11 languages, including English, Spanish, Japanese and Korean |
| FL2VA checkpoint | Text, with zero, one or two images as first and last frames |
| Ref2VA checkpoint | Up to 9 images, 3 videos and 3 audio clips as references, 12 files at most |
The released checkpoints are CFG-distilled, and only the middle of three stages is open.
- 01H3-Context-IR
Hosted. Rewrites your inputs into the structured prompt the model reads.
- 02H3-Base
Open weights. Generates the 768p clip and its audio.
- 03H3-Regenerate-2K
Hosted. Regenerates the clip at 2K. "Not yet open-sourced."
Source: MiniMaxAI/MiniMax-H3 model card, read 13 October 2026
MiniMax calls the first stage "critical to the quality of the final output", so a local run gives 768p output and depends on how well you write the structured prompt yourself. Our AI video generator comparison sets H3 beside other generators.
The license: check this before you download
The license is dated 2 August 2026. Inside the territory it sets these terms:
- Commercial use: allowed, but you need MiniMax's prior written authorization if your commercial products and services make more than 20 million US dollars a year, and a commercial product or service must "prominently display" the MiniMax H3 name in its interface.
- No training other models: you may not use H3 or its outputs "to improve any other artificial intelligence model". H3 clips cannot become a dataset for a different model.
- Disclosure and consent: the acceptable use policy bars publishing machine-generated content "without clearly and prominently disclosing" it, and bars impersonating another person without that person's consent.
- Outputs: "MiniMax claims no rights over the Outputs you generate." You are responsible for them.
Comfy reads the commercial question more strictly. Its H3 guide says commercial use of locally generated outputs "requires a MiniMax commercial license, available through Comfy, the only official reseller", and that generations on Comfy Cloud already include commercial rights. If local output will earn money, ask which reading applies to you. This is not legal advice.
If you are in the US, EU, UK or South Korea
- Apply to MiniMax through its license request form. The license Q&A says the limit means "not yet", not "not ever".
- Buy through Comfy. Its Professional tier starts at $5,000 a month for up to 10 licensed users.
- Use hosted H3. MiniMax describes its API as "globally available with built-in safeguards". The Hailuo app is hosted too, and Comfy Cloud listed its Standard plan at $20 a month.
MiniMax H3 download: weights and ComfyUI files
The original FL2VA and Ref2VA checkpoints are in MiniMaxAI/MiniMax-H3 on Hugging Face. Comfy-Org/MiniMax-H3 holds pruned, quantized versions for ComfyUI, and the templates download those. The text-to-video and image-to-video templates load five files:
| File | Folder under ComfyUI | Size |
|---|---|---|
| minimax_h3_fl2va_pruned_int8_convrot.safetensors | models/diffusion_models | 21 GB |
| qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors | models/text_encoders | 15.7 GB |
| minimax_h3_video_vae_int8_convrot.safetensors | models/vae | 2.81 GB |
| minimax_h3_audio_vae_fp32.safetensors | models/vae | 605 MB |
| minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors | models/loras | 1.96 GB |
That is about 42 GB. The reference template uses the Ref2VA model (another 21 GB) and its own 1.96 GB turbo LoRA. Comfy's notes say to prefer the int8_convrot files "if you are able to use pytorch with cu130" and to use fp8_scaled only if you cannot.
MiniMax H3 VRAM: what the sources say
Neither MiniMax nor Comfy states a minimum. These are the published statements:
| Setup | Source | What it says |
|---|---|---|
| ComfyUI, smallest files | Comfy day-0 post | Footprint cut from 123.6 GB to 42.5 GB. With dynamic VRAM offloading the model can "run locally on a GPU like the RTX 3060" |
| SGLang, one RTX 4090 24 GB | SGLang cookbook | Streams the transformer and text encoder from system memory. "GPU peak stays about 18 GB" |
| vLLM, two 24 or 32 GB cards | vLLM recipe | Layerwise offload on "a 384 GiB-class host" |
Sources: Comfy's day-0 post, the SGLang cookbook and the vLLM recipe. Offloading streams weights from system memory, so a small card needs plenty of RAM and runs more slowly; Comfy gives no timing for the RTX 3060 claim. Without a suitable GPU, see our ComfyUI cloud GPU guide. ComfyUI's own documented requirements are on the ComfyUI statistics page.
How to run MiniMax H3 in ComfyUI
- 1Update ComfyUI
The base templates need 0.30.0 or later. Multiframe Reference needs 0.34.0, Fun ControlNet Union 0.35.0 and FastH3 0.36.0.
- 2Open a template
Template Library, then Video, then a MiniMax H3 workflow. Accept the downloads. No custom nodes are needed.
- 3Set the size
For 16:9 set the Resolution Selector to 0.98 megapixels, which gives 1344x768. The 1.0 step exceeds the pixel cap.
- 4Set the length
A frame count at 24 frames a second. The default is 124, about 5 seconds.
- 5Choose the steps
The default is 20. The turbo LoRA switch drops it to 8, or 4 in reference templates, with slightly lower audio and motion quality.
- 6Write the structured prompt and run
The library ships Text to Video and Image to Video on the FL2VA checkpoint, and Reference to Video, Multiframe Reference and Fun ControlNet Union on Ref2VA, plus FastH3 for 8-step drafts. Details are in Comfy's overview and native workflows page. Two notes matter in practice. A 4-step turbo run can leave a reference "barely applied", so keep 20 steps when a shot must follow a reference. And Sage Attention can "roughly double" generation speed, but it is an optional install that uses the KJNodes pack, which you add through ComfyUI Manager.
For a persona, image to video from a finished still is the most direct route. For Reference to Video, Comfy suggests one character sheet with labeled views.
MiniMax H3 prompting guide
MiniMax publishes the format in two guides, for text and keyframe modes and for reference mode, and Comfy summarizes both in its prompt guide.
- Three fields, in English: integrated_multimodal_description (visuals, action, shots and dialogue in time order), overall_soundscape (1 to 4 sentences) and non_diegetic_music (1 to 3 sentences, or N/A).
- Image to video opens with a fixed line, shown in the example below.
- Time the cuts: "[Shot 2] At 00:03.500, the camera cuts to".
- Dialogue: give each voice an ID such as (S1), and put only the language tag and the exact words inside the tags:
<d>[English] I get off at the next station.</d> - No negative prompt. The templates carry one conditioning input, so "no subtitles" registers text as content. Write "the sign above the door is blank" instead.
- Reference mode: tag inputs in connection order as <Picture 1>, <Video 1> and <Audio 1>, and say what each one controls.
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced. integrated_multimodal_description: [Shot 1] Live-action, cinematic, the young woman shown in <Picture 1> stays at the same cafe window table, keeping her appearance and the soft overcast light. The camera pushes in with small amplitude at slow speed as she lifts her cup, and the woman with a warm, relaxed voice (S1) says: <d>[English] Morning routine, take two.</d> overall_soundscape: Low cafe chatter continues in the background. The ceramic cup clinks once against the saucer. non_diegetic_music: N/A
MiniMax H3 LoRAs
- Turbo LoRAs are built in. The templates load LightX2V's turbo LoRAs for 8-step and 4-step sampling.
- Match the checkpoint. FL2VA and Ref2VA are different weights, and Comfy says to match a LoRA to the build it was made on and to take the pruned build when a repository publishes both.
- Finding them. Hugging Face listed 134 repositories tagged as H3 adapters on 13 October 2026.
- Training. musubi-tuner added experimental H3 LoRA training on 16 September 2026, with a mode that trains from still images. ai-toolkit lists both H3 checkpoints, and DiffSynth-Studio has H3 LoRA training scripts.
A LoRA trained on H3 is a "Model Derivative" under the license: the same territory limits apply, and sharing one requires a copy of the license and a NOTICE file. For dataset basics see our LoRA training guide, and size the run with the free LoRA training steps calculator.
Faces and voices: what you may reference
Reference mode can carry a face from an image and a voice from an audio clip into a new video. Use it for a persona that is a fictional adult you created, or for yourself. Reproducing a real person's face or voice without their written permission is not allowed. The AI influencers library covers persona building and disclosure.
Hosted H3: Hailuo plans and API prices
The Hailuo app runs H3 at 768P or 2K for 4 to 15 seconds. Plan prices are at monthly billing; the estimates of H3 seconds are Hailuo's own.
| Route | Price on 13 October 2026 | What it buys |
|---|---|---|
| Hailuo Free | $0 | One-time trial credits, watermarked downloads |
| Hailuo Standard | $10.50 a month for 1,000 credits | About 143 s at 768P or 84 s at 2K |
| Hailuo Max | $216 a month for 27,000 credits | About 3,857 s at 768P or 2,250 s at 2K |
| MiniMax API, 768P | $0.08 a second | 10 seconds costs $0.80 |
| MiniMax API, 2K | $0.13 a second | 10 seconds costs $1.30 |
| MiniMax API, 768P regenerated to 2K | $0.05 a second | Added to the 768P charge |
On the API, audio references are free, the first 5 reference images are free and then $0.04 each, and a reference video is billed by its length. Hailuo Standard showed a list price of $14.99 beside its discount, and only paid plans carry a commercial-use grant. Every plan is in Hailuo vs Kling, and free tiers are in our free AI video generator comparison. For open models with simpler licenses, see best ComfyUI models.
MiniMax H3: FAQ
What is MiniMax H3?
MiniMax H3 is a video model from MiniMax, launched on 31 July 2026. It takes text, images, video and audio together and generates a clip of 4 to 15 seconds with stereo sound in the same pass. It runs in the Hailuo app, through the MiniMax API, and locally from open weights.
Is MiniMax H3 open source?
The weights are open, but the license is not a standard open-source one. MiniMax published the H3-Base checkpoints on 3 August 2026 under the MiniMax H3 Community License Agreement, which limits where and how they may be used. The prompt-processing stage and the 2K stage were not released.
Where can I download MiniMax H3?
From the official MiniMaxAI/MiniMax-H3 repository on Hugging Face, or as ComfyUI-ready files from Comfy-Org/MiniMax-H3, which the built-in templates fetch for you. The five files the text-to-video template loads add up to about 42 GB.
Can I use MiniMax H3 in the US, UK, EU or South Korea?
Not the open weights under the Community License, which excludes those four territories for the weights and for their outputs. Organizations there can apply to MiniMax for a license or buy one through Comfy, from $5,000 a month. MiniMax says its API is globally available. This is not legal advice.
How much VRAM does MiniMax H3 need?
No minimum is published. Comfy says its smallest files cut the memory footprint from 123.6 GB to 42.5 GB and that, with dynamic VRAM offloading, H3 can run "on a GPU like the RTX 3060".
Does MiniMax H3 support LoRAs?
Yes. ComfyUI loads them, and the official templates ship LightX2V turbo LoRAs. musubi-tuner, ai-toolkit and DiffSynth-Studio list H3 for LoRA training. Match each LoRA to the checkpoint it was made on. The license applies the same territory limits to a LoRA.
How much does MiniMax H3 cost on Hailuo and through the API?
On 13 October 2026 the MiniMax API listed H3 at $0.08 a second at 768P and $0.13 a second at 2K, so a 10-second clip is $0.80 or $1.30. Hailuo Standard was $10.50 a month for 1,000 credits, which Hailuo puts at about 143 seconds at 768P or 84 seconds at 2K.