📦 Run MiniMax H3 Locally: Open Weights in ComfyUI (2026)

The MiniMax H3 weights are real, downloadable, and running in ComfyUI as of this week. What nobody puts in the headline: the repo you download is not the full pipeline behind the Hailuo app's 2K showreels, and the VRAM math deserves a harder look than "runs on an RTX 3060."

Two days after our H3 guide went up saying the weights were "still landing," they landed. This is the follow-up with the local half.

What's actually in the download

The official repo is MiniMaxAI/MiniMax-H3 on Hugging Face, pulled with one line:

hf download MiniMaxAI/MiniMax-H3 --local-dir MiniMax-H3

Inside is H3-Base plus two task-specific BF16 checkpoints in a standard Hugging Face layout: model_index.json, processor, tokenizer, text encoder, transformer, and separate VAEs for visuals and audio. That last part is the tell — the audio path ships in the open release. Local generations come back with the same native stereo track the API produces.

The architecture is a 33B-parameter dense single-stream transformer, with roughly 13B of those parameters sitting in AdaLN-related branches. Keep that 13B figure in mind. It's the reason the optimization story below works at all.

The two pieces MiniMax kept

MiniMax was upfront about this in the release notes, and almost no coverage repeats it: Context-IR is not in the open release, because it depends on a multi-stage hosted workflow. Neither is the 2K regeneration module the hosted product uses to polish output.

So the honest framing is: you get the model, not the product. A local render and a Hailuo-app render of the same prompt will not match, and the gap is widest exactly where the demos impressed you — long-shot consistency and fine 2K detail. I'd rather know that before burning a weekend on downloads. Now you do.

By the numbers

Spec	Local H3 via ComfyUI
Architecture	33B dense, single-stream
Full-precision footprint	123.6 GB
Optimized footprint	42.5 GB (66% smaller)
Quantization	int8 "convrot" + custom kernels
Pruning trick	Modulation weights (~40% of params) replaced by a lookup table
ComfyUI version	0.30.0 or later
Repackaged weights	Comfy-Org/MiniMax-H3
Output	Up to 2K, up to 15 seconds, native stereo
Claimed floor	RTX 3060-class, 12GB VRAM

Sources: the ComfyUI day-0 announcement and MiniMax's release documentation.

The ComfyUI day-0 setup

Comfy shipped support the day the weights dropped, which for a 33B video model is genuinely fast work. Update to 0.30.0, grab the repack from Comfy-Org/MiniMax-H3, and three template workflows appear in the app: video_minimax_h3_t2v.json, video_minimax_h3_i2v.json, and video_minimax_h3_r2v.json. Text-to-video, image-to-video, and reference-to-video with image, video, or audio inputs. First-and-last-frame control made it in too.

Reference-to-video is the workflow to care about. The hosted API's omni-reference system — up to 9 images, 3 videos, 3 audio files — is the best character-consistency input scheme any model currently offers, and r2v is its local cousin.

Can a 12GB card really run this?

Here's how a 123.6GB model shrinks by two thirds. The ComfyUI team found that the modulation weights, about 40% of total parameters, could be pruned and replaced with a functionally equivalent lookup table. The rest gets int8 convrot quantization with custom kernels tuned to cut peak VRAM during inference. Result: 42.5GB, and a stated floor of "a GPU like the RTX 3060."

Read that floor carefully. 42.5GB does not fit in 12GB of VRAM, or in the 16GB on my RTX 4080. The gap gets block-swapped through system RAM, which means the real requirements are a big RAM pool (64GB is the comfortable number) and tolerance for renders that spend half their life shuffling tensors over PCIe. Nobody has published credible seconds-per-clip figures for the 12GB path yet, and until someone does, I'm treating it as a proof of existence, not a workflow.

Quantized community repacks are already multiplying — GGUF and NVFP4 conversions started appearing within 48 hours.

I have not pulled H3 onto my own 4080 yet, and that's a deliberate call: with a 42.5GB footprint against 16GB of VRAM, every generation would be an offloading exercise. For anyone in that VRAM tier who wants to actually evaluate H3 rather than benchmark their patience, renting a 48GB card by the hour costs less than the electricity of a weekend of swapping. My beloved used-3090 advice hits its limit here too — 24GB helps, it doesn't cure.

Should you bother yet?

Worth it now: you own or rent 24GB+, you want dialogue and effects baked into the picture in one pass, or you plan to fine-tune — the whole point of MiniMax shipping BF16 checkpoints. Nothing else you can download today does native audio at this level.

Skip for now: you're on 8–16GB and your work is silent b-roll volume. Wan 2.2 quantized fits your card, renders in minutes, and has a year of ecosystem behind it. H3's local toolchain is a week old. Let other people find the sharp edges — they're finding them enthusiastically as we speak.

How we checked: every figure above traces to the ComfyUI day-0 post, MiniMax's own release notes, or the Hugging Face repos, all read this week. No local benchmark numbers appear because no trustworthy ones exist yet.

The bigger shift: eight months ago the open-weight ceiling was a 14B silent model. Now it's a 33B omni model with stereo audio, and the pruning tricks invented to squeeze it onto consumer cards will outlive H3 itself.