Run MiniMax H3 Image locally: Hugging Face download, SGLang, vLLM-Omni, diffusers

Running the full model is a data-center-class job. The BF16 checkpoint plus its text encoder is roughly 135 GB on disk (68.65 GB + 66.7 GB, per the model card). For a single desktop GPU, the ComfyUI INT8 route is the realistic path.

What you download

Component	Content	Size
`transformer/`	DiT, BF16 (embedding, timestep, output layers FP32)	66.28 GB
`pdd_fused_heads.safetensors`, `pdd_config.json`	Per-step output heads and sampling schedule	12 MB
`vae/`	Image VAE, FP32	2.34 GB
`tokenizer/`, `processor/`	Qwen3-VL tokenizer and processor	23 MB
Text encoder	From `MiniMaxAI/MiniMax-H3`, folder `text_encoder/`	66.7 GB
MiniMax-H3-Image-PE (optional)	Prompt expander, NVFP4	17.9 GB

hf download MiniMaxAI/MiniMax-H3-Image
hf download MiniMaxAI/MiniMax-H3 --include "text_encoder/*"
# optional prompt expander
hf download MiniMaxAI/MiniMax-H3-Image-PE

If you already have MiniMax-H3, the text encoder is not downloaded again.

Architecture and the H3 Image VAE

Text encoder: the full Qwen3-VL text encoder of MiniMax-H3, unchanged; the DiT reads its 50th-layer hidden states.
VAE: a CNN encoder and a 24-layer 2D ViT decoder, 16× spatial downsampling, 24 latent channels. With 2×2 patches, each transformer token covers 32×32 pixels. The repo ships it in FP32 (2.34 GB); ComfyUI's mirror is FP16 (1.17 GB).
DiT: 33.1B parameters, 50 layers, hidden size 5376, 56 attention heads, FFN 14336, 3D multimodal RoPE.

GPU requirements

The model card does not publish a minimum VRAM figure for MiniMax-H3-Image (checked 2026-11-11). It says the vLLM-Omni recipe supports NVIDIA and AMD accelerators, "including NVIDIA H200, RTX PRO 6000 Blackwell, AMD Instinct MI350X and more", and that a single GPU can serve it, with sequence parallelism across several GPUs for lower latency. We do not repeat third-party VRAM numbers here until MiniMax publishes them.

Recommended frameworks

MiniMax recommends SGLang, vLLM-Omni, diffusers and ComfyUI.

SGLang

# one GPU
sglang serve --model-path MiniMaxAI/MiniMax-H3-Image --host 0.0.0.0 --port 30000 --warmup-mode off

# 4 GPUs with Ulysses sequence parallelism
sglang serve --model-path MiniMaxAI/MiniMax-H3-Image --host 0.0.0.0 --port 30000 --warmup-mode off \
  --num-gpus 4 --ulysses-degree 4

Add --text-encoder-path <path> to use a local copy of the MiniMax-H3 text_encoder/ folder.

vLLM-Omni

vllm serve MiniMaxAI/MiniMax-H3-Image --omni --host 0.0.0.0 --port 8000

# 4 GPUs
vllm serve MiniMaxAI/MiniMax-H3-Image --omni --host 0.0.0.0 --port 8000 \
  --num-gpus 4 --usp 4 --text-encoder-tp-size 4

diffusers

# pip install -U diffusers transformers accelerate
import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "MiniMaxAI/MiniMax-H3-Image", dtype=torch.bfloat16, device_map="cuda"
)
image = pipe("Astronaut in a jungle, cold color palette, muted colors, detailed, 8k").images[0]

Request parameters for a local server

Text-to-image is JSON to /v1/images/generations; editing sends references as multipart form data to /v1/images/edits, and their order defines <Picture 1>, <Picture 2> in the prompt. Both return a PNG in data[0].b64_json.

Field	Notes
`prompt`	Required. Used as given, truncated to 4,096 tokens
`size`	Required, `WIDTHxHEIGHT`; output is exactly this size
`seed`, `generator_device`	Seed recommended; `cpu` gives the same noise on every machine
`min_generation_area`	Default 4,194,304 (2048×2048); smaller sizes are generated larger and resized with LANCZOS; 1024 turns it off
`num_inference_steps`	Omit or 6; other values return HTTP 400
`guidance_scale`	Omit or 1; other values return HTTP 400
`negative_prompt`	Not supported

Matching API quality locally

The model expects a detailed English description of the whole image. The API expands short prompts with H3-Image-Context-IR first. Locally, use MiniMax-H3-Image-PE or call the Context-IR endpoint, then pass the result as the prompt. The official examples folder has standard-library scripts for each step (context_ir.py, t2i.py, edit.py, api_generate.py).

Licensing: local use is only licensed inside the "Applicable Territory". See the license summary.