MiniMax H3 Now Generates Video in Eight Steps Inside ComfyUI. The License Excludes the United States.

Alibaba shipped official acceleration adapters on August 26. A ComfyUI node made them loadable five days later. Here is the full workflow, the pruned-checkpoint trap that silently degrades it, and the territory clause in the base model's license that nobody quotes.

Section I.5 of the MiniMax H3 Community License Agreement defines "Excluded Territories" as the European Union, the United Kingdom, the Republic of Korea, and the United States of America.

Section V.4 says you may not use, reproduce, modify, distribute, or display the works "or any of their Outputs or results" outside the Applicable Territory, where Applicable Territory means worldwide minus that list.

MiniMax H3 is currently one of the most widely discussed open-weights video models on the internet. It had 112 fine-tunes on Hugging Face when I checked this morning. It has a native integration in ComfyUI. Alibaba's PAI group published official acceleration adapters for it last week that cut sampling to eight steps, and the adapters have been downloaded roughly 33,000 times. Almost none of the English-language coverage of any of that mentions that, per the license the weights ship under, a large share of the people reading it are not licensed to run the model or to use what comes out of it.

I want to be exact about what I am and am not claiming here, because this is the kind of thing that gets flattened into a scare headline. I am not a lawyer. I am not telling you what enforcement looks like, or whether that clause is enforceable, or what happens to anyone. I am telling you that the clause exists, that it is short and plain and sitting in the LICENSE file, and that if you are about to build a paid workflow on this model it is thirty seconds of reading you have probably not done.

Now let me tell you about the workflow, because the engineering is excellent and the rest of this article is a genuinely useful recipe.

What shipped, and the five-day gap that made it usable

On August 26, alibaba-pai/MiniMax-H3-Acc-LoRAs appeared on Hugging Face. That date is the API's createdAt field, not a listing timestamp, which matters because listing timestamps on Hugging Face are modification times and get people's dates wrong constantly.

The repository contains two files. MiniMax-H3-FL2VA-Acc-8Step.safetensors accelerates the first-and-last-frame and text-to-video paths. MiniMax-H3-Ref2VA-Acc-8Step.safetensors accelerates the reference-image path. Both are rank 64, network alpha 64, BF16, and 1,372,450,680 bytes each. You need one, not both.

The technique is Parallel Decoding Distillation, from an arXiv paper Alibaba cites as 2607.26004. The short version of what it does: instead of the model refining a video over many denoising steps, the adapter carries a set of time-dependent output heads that let several denoising steps' worth of work happen in one network evaluation. There are 32 of those heads, and the loader matches them to whatever sigma schedule your sampler actually produces rather than to a fixed set of positions.

Here is the part that made this a story rather than a footnote. Alibaba shipped it as Python. The quick start says to set model_path and pdd_lora_path inside predict_t2v.py or predict_ref2v.py in the VideoX-Fun repository, then run the script, and it requires diffusers >= 0.40.0. If you write Python and manage your own environments, fine. If your entire creative practice lives in a node graph, which describes most people actually making video with these models, it was inert.

Five days later it was not. (Deno) MiniMax H3 Acc LoRA Loader shipped in a ComfyUI node pack on August 27 and reached its current shape by August 31. It loads Alibaba's official safetensors unconverted, which is the detail I care most about, because the alternative is trusting a community requantization of somebody else's acceleration adapter and hoping the conversion was faithful.

That five-day gap is the pattern worth naming. Research labs ship Python. The distance between a lab release and a creator being able to use it is somebody writing a loader, and that somebody is usually one person with a GitHub account. This node pack has 159 stars.

How the adapter and your scheduler negotiate

The interesting engineering is in what the node does at sampling time rather than at load time.

An acceleration adapter trained for eight steps has, baked into it, an assumption about where the sigma boundaries fall. Sigmas are the noise levels the sampler walks down. Eight steps means eight of them, at particular values. The naive way to build this node would be to force those exact values on you, and the earlier versions of this node did exactly that: they returned their own sampler and their own sigma schedule as node outputs, and every workflow saved against them broke when that changed.

The current design inverts it. The loader returns one thing, a patched model. You build the sampling lane out of stock ComfyUI nodes. At sampling time, the node reads the sigma boundaries your scheduler actually produced and fuses the adapter's 32 output heads to cover those intervals. The adapter adapts to your schedule instead of dictating it.

The practical consequence is that you keep using the sampler nodes you already know, and the pack's documentation can honestly say the loader supports 4 through 12 steps in the Simple scheduler while still telling you that only 8 is trained. Both things are true at once, which is a rarer kind of honesty in this ecosystem than it should be.

Put this into practice

Roughly fifteen minutes if you already have MiniMax H3 running in ComfyUI, plus a 1.37 GB download.

1. Install the node pack. ComfyUI Manager, search Deno Custom Nodes, install, restart. Manual install is git clone https://github.com/Deno2026/comfyui-deno-custom-nodes.git inside custom_nodes, then python -m pip install -r requirements.txt using the same Python that starts ComfyUI.

2. Download the right adapter. Go to alibaba-pai/MiniMax-H3-Acc-LoRAs and take MiniMax-H3-FL2VA-Acc-8Step.safetensors if you do first-and-last-frame or text-to-video work, or MiniMax-H3-Ref2VA-Acc-8Step.safetensors if you drive generation from reference images. Match the adapter to the pipeline. They are not interchangeable.

3. Put the file where the loader looks. Either ComfyUI/models/loras/ or a dedicated ComfyUI/models/minimax_h3_acc_loras/ folder. Configured extra-model roots work too.

4. Add (Deno) MiniMax H3 Acc LoRA Loader to the graph. Connect your native MiniMax H3 diffusion model to its model input. Select the matching adapter in the dropdown. Send the loader's single model output into your normal guider path.

5. Build the sampling lane from stock nodes. BasicScheduler set to simple, steps 8. KSamplerSelect set to euler. Both into SamplerCustomAdvanced. This is the configuration Alibaba trained and the pack recommends.

6. Do not touch three settings. Leave the native MiniMax H3 video sigma shift at 12.0 and the audio sigma shift at 3.0. Leave LoRA strength at 1.0. These are stated defaults, not suggestions, and the trained schedule assumes them.

7. Before you commit disk to any of this, watch the comparisons. The Hugging Face repo hosts result videos in three columns: the un-accelerated MiniMax H3 baseline, the community Minimax-h3-Turbo 4-step checkpoints, and the 8-step Acc-LoRA. Three prompts each for FL2VA at 768p and for Ref2VA. Watch them and decide whether the quality cost is one you will accept, because that is a taste judgment nobody can make for you.

If you want to try it without downloading anything, three community Hugging Face Spaces run these adapters hosted, including mezroui/MiniMax-H3-Acc-LoRAs-demo, which is up. They are user-built rather than Alibaba's, so uptime and queue length are whatever the owner is paying for.

The trap: pruned checkpoints degrade silently

This is the part I would put on a sticky note.

ComfyUI's ecosystem distributes "curve-pruned" MiniMax H3 checkpoints under names containing _pruned_. These exist because the full model is large and people have 12 GB cards. If you are running MiniMax H3 on modest hardware, there is a strong chance a pruned checkpoint is exactly what you downloaded.

The adapter includes 50 AdaLN updates that are full-width. A pruned model's curve basis is 8 wide. They do not fit.

What the loader does about it is clever. It looks for a matching non-pruned MiniMax H3 checkpoint sitting in models/diffusion_models/, reads only that checkpoint's small FP32 time-embedder section, and uses it to derive an in-memory bridge that rebases all 50 AdaLN updates onto the pruned 8-wide curve. It never loads the full checkpoint into memory to do this. That is a genuinely nice piece of engineering.

And if you do not have that full checkpoint installed, it warns once, skips those 50 updates, and keeps going in a documented compatibility mode.

Read that again. It keeps going. You get video. It just is not the video the adapter was supposed to produce, and the only signal is one console warning that scrolled past while you were looking at the queue. If you run pruned checkpoints because you are short on disk in the first place, the fix is to also keep a full checkpoint on disk, which is the opposite of why you pruned.

There is a second, smaller trap. Workflows saved against the older v0.7.92 through v0.7.94 three-output version of this loader get migrated automatically when you open them on the ComfyUI canvas, with the old sampler and sigmas links moved onto editable stock nodes. You must save the workflow once after it opens for that to stick. Raw API prompt JSON does not run the frontend migration at all and has to be exported again from a migrated UI workflow. If you already saved a file after its sampler links vanished, you are reconnecting those by hand.

Honest limitations

The license is the big one, and it is not the usual open-weights caveat. Most restrictive model licenses limit commercial revenue or forbid training other models. This one does both of those, and adds a geography. Applicable Territory is worldwide excluding the EU, the UK, South Korea and the US. Section V.4 extends the restriction to Outputs, not just the weights. Separately, Section IV.1 requires prior written authorization from MiniMax if your commercial products and services generate more than 20 million US dollars a year, and Section IV.2 requires you to prominently display "MiniMax H3" in the interface of any commercial product or service using it. Governing law is Hong Kong SAR. The license invites people in excluded territories to contact MiniMax about obtaining a license, so the door is not shut, it is just not open by default.

The adapters inherit that license. Alibaba's repository card sets license: other with license_name: minimax-h3-community-license-agreement and points at MiniMax's LICENSE file. Alibaba's own contribution here is not separately Apache or MIT. The ComfyUI node pack is GPL-3.0-only, which covers the node code and does not change the weights' terms, and its README says so directly.

Eight steps is the only trained configuration. The pack lets you pick 4 through 12 and describes other descending schedules and split-sigma latent-upscale passes as "available for experimentation rather than guaranteed quality improvements." That is the maintainer being careful. Treat anything other than 8 as an experiment you are running, not a setting you are choosing.

The quality comparison is Alibaba's own. The side-by-side videos are the lab's chosen prompts, generated by the lab, hosted by the lab. They also note the test cases come from the Minimax-H3-Turbo examples and were generated at LoRA weight 1.0 at both 4 and 8 NFE. This is more transparent than most acceleration claims and it is still the author's own evaluation.

There is no VRAM figure anywhere in this stack. Not on Alibaba's card, not in the node pack's section for this loader. What you need is whatever native MiniMax H3 needs on your machine, and eight steps reduces your time, not your memory ceiling. Fewer steps does not make a model fit.

What I would actually do with this

If you are somewhere the license covers, this is close to a free speedup on a model you already run, using the lab's own file, with a loader that does not lock you into its opinions about samplers. Install it, keep 8 steps, keep a full checkpoint on disk if you use pruned ones, and get on with making things.

If you are in the United States, the EU, the UK or South Korea, the honest answer is that you have a decision to make that a tutorial cannot make for you, and the first step is reading four short paragraphs of a LICENSE file rather than taking my summary of them.

The larger thing I keep turning over is this. The open-weights video ecosystem has spent two years training people to read one word, the license name on the model card badge, and stop. "Apache." "MIT." "Community license." That last one is doing an enormous amount of hidden work. It has come to mean "some restrictions, probably a revenue cap, probably fine for me." MiniMax H3's community license means something quite different from Llama's, which means something different again from Krea's, and the only way to know is to open the file.

I would rather that be a boring habit than a scandal. Open the LICENSE. It takes a minute. If you have found a territory clause in a model you were already shipping work with, I would like to hear about it, because I suspect this is not the only one.

---

Medium metadata

Suggested kicker image: a ComfyUI node graph showing the Acc LoRA Loader feeding a BasicScheduler at simple/8 into KSamplerSelect at euler.

Sources:

• alibaba-pai/MiniMax-H3-Acc-LoRAs model card
• MiniMax H3 Community License Agreement
• Deno Custom Nodes README and CHANGELOG
• VideoX-Fun, where the official predict scripts live
• Parallel Decoding Distillation, arXiv 2607.26004