Local deployment · verified 2026-11-13
Running MiniMax H3-Image locally, and the 135 GB question
Short answer
The repository says 68.65 GB, and that number is true only for the diffusion transformer. The full H3-Image stack also pulls a 66.7 GB text encoder from the base MiniMax-H3 repository. Budget roughly 135 GB of model files before runtime state, plus 17.9 GB if you want the local prompt expander. Five runtimes are documented; none comes with one universal minimum-VRAM number.
Weights under the MiniMax-H3-Image Community License Agreement Model card and commands →
- 135 GBmodel filesDiT plus text encoder
- 5documented runtimesSGLang, vLLM-Omni, diffusers, LightX2V, ComfyUI
- 6sampling stepsno classifier-free guidance
- 4 GPUsdocumented scale-outUlysses sequence parallel
Before the hardware decision
- The weights licence is not worldwide. The EU, UK, South Korea and US are excluded from the Applicable Territory.
- Hosted use needs extra permission. A service that makes the model available to third parties requires separate written authorisation.
- This site has not run local H3-Image inference. The commands and file sizes below are from MiniMax and Hugging Face; they are not a measured build from our machines.
What you actually download
The H3-Image repository is not self-contained. The pipeline loads its text encoder from the MiniMax-H3 repository, which is why a machine that already has H3 does not download that half twice. A fresh install does.
| Field | Value |
|---|---|
| BF16; 66.28 GB inside the repository, plus FP32 embedding, timestep and output layers. | |
| Separate download from MiniMaxAI/MiniMax-H3 under text_encoder/. | |
| FP32, included in the H3-Image repository. | |
| The same Qwen3-VL processor used by MiniMax-H3. | |
| pdd_fused_heads.safetensors and pdd_config.json. | |
| NVFP4 MiniMax-H3-Image-PE, used for local text-to-image prompt enhancement. |
The model-card download commands are:
hf download MiniMaxAI/MiniMax-H3-Image
hf download MiniMaxAI/MiniMax-H3 --include "text_encoder/*"
# optional prompt expander
hf download MiniMaxAI/MiniMax-H3-Image-PE Hugging Face model card, read 2026-11-13 Prompt expander card →
Five runtimes, and what each is for
Choose by the interface you need rather than by a leaderboard. SGLang and vLLM-Omni give you an HTTP service with multi-GPU options. diffusers is the direct Python path. ComfyUI is the visual workflow route. LightX2V is the other serving path MiniMax links.
| Runtime | Role | What the model card provides |
|---|---|---|
| SGLang | Serving | The model card's single-GPU example and a four-GPU Ulysses configuration. |
| vLLM-Omni | Serving | Per-hardware recipes are published for NVIDIA and AMD accelerators. |
| diffusers | Python | The shortest path from a prompt to an image; less focused on production serving. |
| LightX2V | Serving | MiniMax links its own usage guide from the model card. |
| ComfyUI | Visual workflow | The model card links separate text-to-image and edit templates. |
A one-GPU server, and the four-GPU split
The model card's SGLang example starts a server with no warm-up and leaves the text encoder in the Hugging Face cache. The four-GPU variant splits each request with Ulysses sequence parallelism, which is the pattern to investigate when a single device is short on memory or too slow.
# one GPU
sglang serve --model-path MiniMaxAI/MiniMax-H3-Image \
--host 0.0.0.0 --port 30000 --warmup-mode off
# four GPUs
sglang serve --model-path MiniMaxAI/MiniMax-H3-Image \
--host 0.0.0.0 --port 30000 --warmup-mode off \
--num-gpus 4 --ulysses-degree 4
The vLLM-Omni recipe follows the same shape: one GPU by default, or four GPUs with
--num-gpus 4 --usp 4 --text-encoder-tp-size 4. MiniMax names the H200, RTX PRO
6000 Blackwell and AMD Instinct MI350X among the accelerators in its per-hardware recipes.
That is a list of tested examples, not a claim that any particular 24 GB card is enough.
Model card deployment commands, read 2026-11-13 No independent runtime benchmark on this page
The ComfyUI route
MiniMax publishes separate text-to-image and edit templates for ComfyUI. The text-to-image
template receives a prompt, a size and a seed; the edit template also needs the reference
images in upload order. Keep the six-step and guidance-free settings unless the template
has been built for a different distilled schedule. The local API rejects other values for
num_inference_steps and guidance_scale.
For prompt expansion inside ComfyUI, the optional PE checkpoint is a separate model. The hosted API applies H3-Image-Context-IR automatically; a local install must either run PE itself or send an already detailed prompt.
The hardware arithmetic, stated honestly
The weights fit is the part you can calculate: 68.65 GB for the DiT and 66.7 GB for the text encoder. The rest depends on the runtime, precision of activations, sequence length, attention implementation and whether the text encoder is co-resident. That is why this page does not turn a file size into a minimum-VRAM claim.
The sensible test is a single request at the resolution you intend to sell, on the runtime you intend to operate, with the text encoder loaded the same way. If the process fits but the latency is wrong, the four-GPU path is the documented next step. If the process does not fit, a smaller quantisation may exist, but it is a community artefact rather than the published BF16 reference and should be treated as such.
The licence check comes before the GPU check
The MiniMax-H3-Image Community License Agreement excludes the European Union, the United Kingdom, South Korea and the United States from the Applicable Territory. It also requires separate written authorisation for hosted services and for organisations above $1 million in annualised revenue. A local benchmark run in one of those markets is not authorised merely because the files can be downloaded.
Licence read 2026-11-13 Read the agreement →