MiniMax H3 Model (CRT AutoDL)

MiniMax H3 Model (CRT AutoDL) is the diffusion-model half of CRT's MiniMax H3 stack. H3 is MiniMax's 33B omni-modal video model - text, image, video, and audio in one context, generating clips with native stereo audio instead of bolting a sound pass onto a silent generator. It landed open-weights in August 2026 and immediately became one of the biggest local-video deals of the year, and this node is the "just give me the model" entry point: pick a variant, and it downloads the weights on first use into ComfyUI/models/diffusion_models/, then hands you a MODEL.

The H3 reality check (read this first)

H3 is genuinely excellent - it hit #1 in the video-with-audio arena and the community's launch reaction was unusually strong. But there's a catch that's easy to miss and impossible to ignore: the MiniMax H3 Community License excludes the European Union, the UK, the Republic of Korea, and the United States from its applicable territory. If you're in one of those regions, you're not licensed to run the local weights at all, regardless of what the node will happily download. The hosted Hailuo API is separate and global; only the weights carry the geography restriction. This node doesn't (and can't) check your location - it just downloads. Read the license before you run it.

How it works

model_name gives you two model families, each with a standard and a "Light" quant:

• FL2VA - the flagship text/video-to-video variant. Also the H3 analog of LTX-style generation from text or a reference clip.
• REF2VA - the reference/reconstruction-oriented variant, the one that leans on input context. (In H3's ComfyUI ecosystem, "2VA" families select which conditioning route the model is built around.)
• Both come as pruned_int8 (the Comfy-Org repacks) and as Light W4A8 (Kijai's experimental w4a8 quants). The Light versions are the memory play - w4a8 means 4-bit weights, 8-bit activations, and for a 33B model that's the difference between "fits" and "page to system RAM."

Like every AutoDL diffusion loader, you also get the CRT fast-settings trio: patch_cublaslinear (default off), attention_method (default disabled; the menu has PyTorch plus a shelf of SageAttention modes for when you've installed SageAttention and want the speed), and enable_fp16_accumulation (default on, matching community practice for H3).

Output is a single MODEL, which feeds your sampler. H3 is a full stack, so you'll also need the sibling nodes - MiniMax H3 CLIP (a Qwen3-VL 32B text encoder, INT8 or NVFP4), MiniMax H3 VIDEO VAE, and for audio output MiniMax H3 AUDIO VAE. CRT ships all of them as AutoDL nodes, so the whole pipeline installs itself.

Installing it

Standard CRT install - once for the pack, not per node:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r requirements.txt

or ComfyUI Manager → search CRT-Nodes, then restart. The requirements are heavy (opencv, transformers, faster-whisper, librosa…), so budget install time. The model download is the bigger wait: the INT8 H3 checkpoints are tens of gigabytes, so a 1 GB/s connection is your friend.

Common issues

• First run looks hung - it's the multi-GB download. Watch the console for the progress bar.
• Don't mix variant families - an FL2VA model with the REF2VA CLIP setup will misbehave. Keep model, CLIP, and VAEs from the same family.
• OOM on a 16 GB card - drop to a Light W4A8 model and the NVFP4 CLIP. H3 full-precision is not a consumer-VRAM proposition.
• License, license, license - see above. This is the one CRT AutoDL node where you should check your region before clicking run.
• Update breakage - red nodes/NaN → right-click → Fix node (recreate).