AI & Creative Tools

A 56GB Image Model Just Got Small Enough to Run Natively Inside ComfyUI

MiniMax's H3 image model is a serious piece of hardware to run on its own. Pruna's quantized P-H3-Image builds, now natively supported in ComfyUI with no custom nodes, bring it down to a size regular creative workstations can actually load.

The pattern of “powerful new image model ships, then promptly requires more VRAM than almost anyone outside a research lab actually has” repeats often enough that quantization work has become its own beat worth watching. P-H3-Image, Pruna AI’s quantized build of MiniMax’s H3 image model, got native ComfyUI support in early November — merged without requiring any custom nodes — which is the part that actually changes who can use it.

What’s actually different between the builds

Pruna’s repo ships the model in three tiers, and the differences aren’t cosmetic:

  • bf16, the full-precision set, is a roughly 55GB download and peaks at about 56GB of VRAM with the prompt upsampler loaded — squarely enthusiast-workstation-or-better territory, not a casual download.
  • int8 applies a grouped Hadamard rotation against outliers and quantizes activations on the fly, which Pruna describes as staying “very close” to the full-precision output.
  • w4a8 pushes further, putting weights on a 4-bit, 16-level codebook — Pruna’s own characterization is blunter here: this tier “changes more of the details,” the expected tradeoff of a more aggressive compression pass.

That spread matters because it turns one monolithic “can you run this or not” decision into a genuine choice: take the full-fidelity hit on hardware that can afford it, or trade some detail for a model that actually fits a serious hobbyist GPU rather than a datacenter card.

Why “no custom nodes” is the real headline

A quantized community build existing is routine at this point — what’s less routine is ComfyUI merging native support for it (tracked as PR 17506 in the project’s own issue system) rather than leaving it to a third-party custom node that has to be separately installed, kept updated, and trusted. Native support means the model shows up in a standard ComfyUI workflow the same way any first-party node would, lowering the setup bar from “find and vet a community extension” to “download the files and go.” A separate GGUF build from a third party, AtomicChat, extends that same access further to stable-diffusion.cpp-style setups via ComfyUI-GGUF’s Unet Loader.

The licensing fine print

P-H3-Image’s weights are derived from MiniMax-H3 and ship under the MiniMax H3 Community License — free for the use cases that license permits, but commercial use requires requesting a separate license through Comfy, a detail easy to miss when a model shows up framed purely as “now runs locally,” and worth checking before building a paid product on top of it.