Local deployment guide

How to Run MiniMax H3 Locally: The Complete Beginner's Guide

Set up MiniMax H3 open weights in ComfyUI, understand the model files and VRAM trade-offs, and make a first local 768p video with native stereo audio.

MiniMax launched MiniMax H3 on July 31, 2026, then released the H3 model weights on August 3, 2026. The open-weight release makes it possible to run H3-Base locally for 768p video generation with native stereo audio using supported frameworks such as ComfyUI, without per-generation API fees. It does not remove the cost or complexity of storage, hardware, electricity, or cloud GPU time.

Read the license before downloading

The MiniMax H3 Community License defines an Applicable Territory. The European Union, United Kingdom, Republic of Korea, and United States of America are excluded territories. Users in excluded territories should review the current license and contact MiniMax regarding separate authorization before deploying the open weights. The hosted MiniMax API is a separate service governed by its own terms; this does not make it an automatic compliance option. This guide is informational and is not legal advice.

Open MiniMax H3 Studio

Local setup is not the quickest path for every project

Local H3 is useful when you need to study or customize the open-weight workflow and can operate the required hardware. If you want browser-based generation or a 2K workflow without downloading large checkpoints and configuring a GPU environment, use the MiniMax H3 AI Video Generator instead.

What Exactly Did MiniMax Release?

MiniMax H3 has three connected layers. They are not interchangeable downloads, and only one is the main local-generation target at this time.

LayerWhat it doesLocal status
H3-Context-IRA hosted preprocessing and orchestration system that turns rich multimodal inputs into context for generation.Not part of the open-weight release.
H3-BaseThe local-generation core. Its H3 Omni Transformer has 33B parameters, with roughly 13B in AdaLN-related branches that can be precomputed or cached for inference-only deployment.The main local workflow, centered on a 768px short edge.
H3-Regenerate-2KRegenerates an H3-Base 768p output together with the original context to reach 2K; it is not a conventional super-resolution upscaler.Not released as open weights at publication; MiniMax has said it plans to release it after stabilization.

FL2VA vs Ref2VA: Choose the Right Local Workflow

FL2VA

FL2VA covers Text-to-Audio-Video, first-frame-to-video, last-frame-to-video, and first-plus-last-frame-to-video. In ComfyUI, use MiniMaxH3ImageToVideo for image-driven generation and optional first/last-frame control.

Ref2VA

Ref2VA is multimodal reference generation: reference images can guide identity or appearance, reference video can guide motion or camera behavior, and reference audio can guide voice, sound, or rhythm. Current official limits allow up to 9 reference images, up to 3 reference video clips, and up to 3 audio clips, with a maximum of 12 supported reference files. Prepare video and audio clips within the duration restrictions in the current official workflow documentation.

Native nodes also include EmptyMiniMaxH3LatentAV, MiniMaxH3ReferenceToVideo, and MiniMaxH3SigmaShift. The template library’s three main examples are Text-to-Video, Image-to-Video, and Reference-to-Video; first/last-frame control belongs to the image-to-video nodes rather than a separate set of local templates.

MiniMax H3 Hardware and VRAM Requirements

There is no single official MiniMax or ComfyUI minimum-GPU promise for every H3 workflow, quantization, driver, or operating system. Treat the ranges below as planning guidance, not a compatibility guarantee.

HardwareWhat to expect
12 GB VRAMPossible with aggressive CPU or system-memory offloading and reduced-memory variants according to community testing. Expect slow loading and generation; this is not an official minimum requirement.
16–24 GB VRAMMore practical, although model components may still need offloading depending on the selected workflow and quantization.
32 GB+ VRAMMore headroom for reduced-memory local workflows, but the complete H3 stack remains large and storage-heavy.
AMD or Apple SiliconNot automatically impossible. Official ComfyUI H3 guidance currently centers on the native workflow, while NVIDIA/CUDA is the most documented path; alternative and community support changes quickly.

The recommended FL2VA reduced-memory files are still substantial: the diffusion model is roughly 21 GB, the text encoder roughly 15.7 GB, and the video and audio VAEs roughly 5.21 GB and 605 MB. Plan for about 42.5 GB of model downloads before workflow files and outputs.

Step 1: Install or Update ComfyUI

Update ComfyUI to ComfyUI 0.30.0 or later. The beginner-friendly path is the native Template Library: open Video, choose a MiniMax H3 Text-to-Video, Image-to-Video, or Reference-to-Video template, and follow the model-download prompt. This avoids accidentally downloading every variant in the model repository.

Use the native workflow first. It gives you the supported node wiring, model locations, and resolution controls before you begin making custom changes.

Step 2: Download Only the MiniMax H3 Model Files You Need

Do not use a whole-repository download. The Comfy-Org repository contains many variants and is roughly 465 GB in total. For a manual FL2VA setup, use the current official file set below and replace ComfyUI/models with the path to your own installation; Windows and Linux paths can differ.

hf download Comfy-Org/MiniMax-H3 diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors vae/minimax_h3_audio_vae_fp32.safetensors vae/minimax_h3_video_vae_fp16.safetensors --local-dir "ComfyUI/models"

When your PyTorch/CUDA environment supports it, Comfy-Org currently recommends minimax_h3_fl2va_pruned_int8_convrot.safetensors. Otherwise, choose the fp8_scaled variant only as described in the official repository. Do not treat third-party community NVFP4 transformers as the official Comfy-Org download.

diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors

text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

vae/minimax_h3_audio_vae_fp32.safetensors

vae/minimax_h3_video_vae_fp16.safetensors

Step 3: Load the Official ComfyUI Workflow

In Template Library, open the native T2V, I2V, or R2V MiniMax H3 workflow. Confirm that the workflow finds the diffusion model under models/diffusion_models, the Qwen text encoder under models/text_encoders, and both VAEs under models/vae.

For full-quality local H3-Base output, use the workflow’s Resolution Selector at about 1.0 megapixel for 16:9, which is roughly 1344×768. Keep dimensions on the H3 resolution grid: ComfyUI documents a 768px short edge, capped at 768×1344 and rounded to multiples of 32.

Step 4: Generate Your First MiniMax H3 Video

  1. Start with the Text-to-Video template to confirm the stack works before adding references.
  2. Write one clear scene, its motion, one camera idea, and its sound direction in the same prompt block.
  3. Use the template preview resolution first, then raise the megapixel setting once the workflow completes reliably.
  4. For image or reference workflows, add one asset at a time and give each input a clear role.

For prompt structure and examples, see the site's MiniMax H3 Prompt Guide. For a browser-first creative workflow, see How to Use MiniMax H3.

Troubleshooting a Local MiniMax H3 Setup

Out-of-memory errors

Lower the preview resolution, use the supported reduced-memory model variant, and enable the available offloading options. Do not assume an OOM result means the model is incompatible; memory behavior varies with the selected workflow and environment.

Missing model errors

Recheck the four folder locations exactly. The diffusion model, text encoder, video VAE, and audio VAE have separate destinations.

Slow startup or generation

Large checkpoints can take time to load, especially with offloading. Verify storage throughput, available system memory, and the active device before changing model files.

GGUF downloads

MiniMax has not released an official GGUF build. Community GGUF conversions now exist on Hugging Face, but they are third-party releases rather than official MiniMax checkpoints. Verify the repository, workflow compatibility, model provenance, and MiniMax H3 Community License before downloading.

Local vs Online MiniMax H3

Choose local H3-Base when…Choose the online workflow when…
You need open-weight experimentation, can manage large model files, and have a compatible local or rented GPU environment.You need 2K output without configuring a local GPU workflow, downloading checkpoints, or managing model compatibility.
You are prepared to review the Community License and deployment requirements for your territory.You want to create in the browser under the hosted service's separate terms and current product availability.

MiniMax H3 Local Install FAQ

Can I run MiniMax H3 locally?

MiniMax released H3-Base open weights, and ComfyUI documents native local workflows for text-to-video, image-to-video, and reference-to-video. Running locally still requires substantial model downloads and compatible compute.

Is MiniMax H3 local generation free?

Local use can avoid per-generation API fees, but it still has hardware, electricity, storage, and possibly cloud-GPU costs. The applicable MiniMax H3 Community License also matters.

What is the official local MiniMax H3 format?

This guide uses the official Comfy-Org repackaged model files and ComfyUI’s native H3 workflows. MiniMax has not released an official GGUF build; community GGUF conversions on Hugging Face are third-party releases, so verify provenance, compatibility, and licensing before downloading.

Can local ComfyUI produce full 2K MiniMax H3 output?

H3-Base is documented around a 768px short-edge native canvas. Full 2K uses H3-Regenerate-2K, which combines the base output with the original context; it was not part of the open-weight release at this article’s publication.

Which MiniMax H3 workflows are available locally?

ComfyUI’s template library documents three native examples: Text-to-Video, Image-to-Video, and Reference-to-Video. The Image-to-Video nodes also support first- and last-frame control.

Sources

Prefer to Generate Without a Local Setup?

Use MiniMax H3 in the browser when you want 2K output without downloading model files or configuring a local GPU workflow.

Open MiniMax H3 Studio