How to run MiniMax H3 locally.

What it makes.

H3 is MiniMax’s open video model, on Hugging Face since late July 2026. It generates the picture and a 32 kHz stereo soundtrack together: voices, sound effects and music, all from one prompt.

• Length: 5 to 15 seconds at 24 fps. The model card says 4 to 15. Frame counts snap to steps of 17, plus 5: 124 frames is five seconds.
• Size: 768p, which is 1344 × 768 at 16:9. The 2K upscale MiniMax shows off runs only on their API.
• Two checkpoints: FL2VA does text to video, or animates from a first and last frame. Ref2VA takes references: up to nine images, three videos and three audio clips.
• No CFG, no negative prompt. The released weights are guidance-distilled, so one pass per step.

Under the hood it is a 33B transformer, but about 13B of that sits in branches whose outputs can be computed ahead of time. The pruned files leave those out, which is why they are 21 GB and not 34.

Files you need.

All four from Comfy-Org’s repackage. These are the ones Comfy’s template uses.

• Model minimax_h3_fl2va_pruned_int8_convrot.safetensors ComfyUI/models/diffusion_models/
  21.0 GB Download
• Text encoder qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors ComfyUI/models/text_encoders/ doesn’t need a Blackwell GPU
  15.7 GB Download
• Video VAE minimax_h3_video_vae_int8_convrot.safetensors ComfyUI/models/vae/ fp16: 5.2 GB
  2.8 GB Download
• Audio VAE minimax_h3_audio_vae_fp32.safetensors ComfyUI/models/vae/
  0.6 GB Download
• Turbo LoRA minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors ComfyUI/models/loras/ optional, 8 steps. A 4-step 768p LoRA is in the Comfy-Org repo
  2.0 GB Download

models/
├── diffusion_models/
│   └── minimax_h3_fl2va_pruned_int8_convrot.safetensors
├── loras/
│   └── minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors
├── text_encoders/
│   └── qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
└── vae/
    ├── minimax_h3_video_vae_int8_convrot.safetensors
    └── minimax_h3_audio_vae_fp32.safetensors

Other builds

Comfy-Org’s readme says to prefer int8_convrot if your PyTorch is built for CUDA 13.0 (cu130), and fp8 only if int8 won’t run. Ref2VA comes in the same sizes.

Model sizes are the pruned FL2VA files unless the row says unpruned.

What fits your computer.

H3 streams its weights between system RAM and the GPU, so RAM matters as much as VRAM. The diffusers docs expect around 75 GB of host RAM at int8 on a 12 to 16 GB card.

• 8 GBBarely

  No 8 GB card report. One 16 GB owner capped ComfyUI at 8 GB and still ran 640 × 384, streaming about 270 GB from disk per run.

• 12 GBTight

  Small canvases like 960 × 544, with the weights in system RAM. An RTX 4070 hung at “Model Initializing”, a known intermittent issue.

• 16 GBOffloads

  RTX 5070 Ti with 32 GB of RAM: 640 × 384, 226 frames, with --disable-pinned-memory and --reserve-vram 1.5.

• 24 GBFits

  RTX 4090 with 62 GB of RAM: 1344 × 768, five seconds, with SageAttention. VRAM peaked at 15.7 GB.

• 32 GBFits

  RTX 5090. One owner measured Turbo at 8 steps 2.2 times faster than 20 steps; clips of 345 frames and more stalled.

• Mac 48 GB+Use h3.c

  No verified ComfyUI run on a Mac. antirez’s h3.c runs H3 on Metal, peaking at 27.8 GB (int8) to 39.1 GB (bf16) at 512 × 512.

• 16 GB cards
• RTX 5060 Ti 16 GB
• 24 GB cards
• RTX 5090
• Mac with Apple Silicon

Set it up.

1. 1

   Update ComfyUI

   H3 needs ComfyUI 0.30.0 or later. Desktop follows stable releases, so it can lag behind. In a manual install:

   Terminal, in the ComfyUI folder Copy

   git pull
   pip install -r requirements.txt

2. 2

   Download the files

   Model, text encoder and both VAEs from the list above, about 40 GB. Add a Turbo LoRA if you want 8-step clips.

3. 3

   Put them in their folders

   Model in diffusion_models, encoder in text_encoders, both VAEs in vae, LoRA in loras. Restart ComfyUI.

4. 4

   Open the template

   In the template browser, pick MiniMax H3: Text to Video, MiniMax H3: Image to Video or MiniMax H3: Reference to Video. The image template does text to video too when no picture is connected.

5. 5

   Check the settings

   Load CLIP uses type minimax. There are two VAE loaders, one for video and one for audio. Set the duration in seconds; a math node turns it into a valid frame count. Keep the Resolution Selector’s multiple at 32.

6. 6

   Write a structured prompt

   Describe the shots, camera moves and sound in one block, in the labelled sections from MiniMax’s prompting guide. Freeform prompts follow poorly. MiniMax’s own prompt rewriter, which the card calls critical to quality, is only on their API.

Settings that work.

Steps

20

CFG

noneBasicGuider, no negative

Sampler

res_multistep

Scheduler

simple

Shift

12ComfyUI’s default for H3

Size

1344 × 768768 px short edge

Length

5 to 15 smodel card: 4 to 15 s

Frames

124 for 5 ssteps of 17, plus 5

With a Turbo LoRA

8-step LoRA

8 stepsstrength 1.0

4-step LoRA

4 stepsmade for 768p

CFG

none

Sampler

res_multistep

Match the steps to the LoRA. In one RTX 4090 test, 4 steps looked mushy, 8 were coherent and usable, and 12 was the tester’s pick for everyday use.

Sizes from Comfy’s template note, at a multiple of 32. Avoid 1.0 MP: 1376 × 768 goes over H3’s pixel cap.

How fast.

The RTX 4090 ran ComfyUI 0.30.0 with SageAttention and 62 GB of RAM. The M5 Max figure is from h3.c, a separate Metal program, not ComfyUI.

When it goes wrong.

It hangs at “Model Initializing...”

A known, intermittent issue with dynamic VRAM. Remove MultiGPU and other custom nodes that touch model loading, and try again.

CUDA out of memory on 24 GB

Keep dynamic VRAM on, which is the default, and use a smaller canvas or fewer frames.

Windows crashes with an access violation in the video VAE decode

Start ComfyUI with --disable-async-offload --disable-pinned-memory.

The PC swaps heavily with 32 GB of RAM

Add --disable-pinned-memory.

to() received an invalid combination of arguments - got (NestedTensor…)

VAE Decode (Tiled) doesn’t work with H3. Use the plain VAE Decode.

shape '[25600, 2560]' is invalid for input of size 3930620

Load CLIP failing on the encoder, most likely an older ComfyUI without nvfp4 support. Update, and check the type is minimax.

Static or popping in the audio

Update ComfyUI. For one user, a fresh portable 0.30.0 fixed it.

Garbled speech in the first moment

It depends on the seed. One user fixed it by writing dialogue in plain quotes instead of the <d> tags.

Mushy video with the Turbo LoRA

4 steps is too few for most clips. Use the 8-step LoRA at 8 steps.

The GGUF file won’t load

ComfyUI-GGUF doesn’t support H3 yet. Use the int8 safetensors files.

Questions.

Sources: MiniMax H3 model card, licence, licence Q&A thread, ComfyUI MiniMax H3 tutorial, diffusers H3 docs, Comfy-Org repackage, RTX 4090 test [1], 16 GB notes [2], RTX 5090 notes, h3.c [3], ComfyUI issue #15628, #15337, #15274, #15614.