MiniMax H3-Image Local Setup: 135 GB and ComfyUI

H3-Image's weights are public, and local deployment is an officially supported path — but "download 68.65 GB and run it" is a misunderstanding. A first deployment actually needs about 135 GB of downloads, split across two repositories; there are five frameworks, with SGLang officially recommended; and ComfyUI shipped native support on launch day. The biggest gap: MiniMax still hasn't published any VRAM guidance.

The file list: about 135 GB across two repos

The text encoder is not in the H3-Image repo — it must be downloaded separately from the MiniMax-H3 repo, and it's the easiest step to miss. Users who already run H3 locally don't need to re-download it; the encoder is identical. Note that TechToHeart quotes about 127 GB of disk for the local package, which differs from 135 GB (possibly compressed or a different counting basis); go by the official file sizes. PE is an optional prompt-expansion model (17.9 GB NVFP4, fine-tuned from Qwen3.8-27B) for reproducing the API's prompt expansion locally — the model runs without it.

Framework choice: SGLang is the official recommendation

The model card recommends SGLang as the reference path, with a cookbook and a single-GPU command, supporting --ulysses-degree sequence parallelism. vLLM-Omni provides recipes for H200, RTX PRO 6000 Blackwell and MI350X hardware, but its day-one recipe only exercises 9 reference images per request — 16 is the model limit, 9 is what that framework's recipe measured. diffusers and LightX2V are also available. The official GitHub repo is MiniMax-AI/MiniMax-H3-Image, and ModelScope hosts a mirror.

Practical advice: to get running fastest and follow the official cookbook, use SGLang; if you already have a vLLM-Omni setup on H200 or Blackwell cards, use vLLM-Omni; if you want a node-based graphical interface, use ComfyUI (next section). diffusers suits those who want to write their own inference code.

ComfyUI: native support on day one, two ways to run it

The ComfyUI blog announced native support on November 10 at 16:05 UTC, with v0.45.0 shipping the same day, including H3-Image nodes and official text-to-image and editing workflow templates. Text-to-image and editing share one graph: connecting reference images switches to editing mode. Two ways to run it: download the weights and run on your own GPU, or use Partner Nodes to call the MiniMax API without downloading weights. The Qwen3-VL text encoder is shared with H3, so existing H3 users don't re-download it.

Partner Nodes is a notable middle path: no 135 GB download, no GPU of your own, yet you can call the official API from inside the ComfyUI interface. The trade-off is API billing, with prompt expansion and similar behavior handled API-side — a good fit for trying the results before deciding on local deployment.

The missing VRAM guidance

MiniMax publishes only file sizes, with no VRAM guidance at all: whether a single card can run it, how much memory it needs, whether quantization options exist — none of it is stated. File size is not VRAM — inference also needs KV cache, activations and intermediate results, and while six sampling steps is few, the latent for a 4K generation is itself large. Don't estimate graphics cards from 135 GB, and don't estimate them from 68.65 GB either. The third-party distilled P-H3-Image has a 4-bit path at about 21 GB, but that's a different model and doesn't apply to the official H3-Image. This site will update when official or community VRAM data becomes available.

Read the license before deploying

The weights license excludes the US, EU, UK and South Korea — anyone there needs a separate license from MiniMax before using the weights; written authorization is required above $1M annualized revenue or for hosted services; commercial interfaces must display the model name; outputs may not train other models. Read the full LICENSE before downloading; see our license deep dive.