MiniMax H3 Singularity V2: What’s New, Lighter Files and License

MiniMax H3 Singularity is one of those names that keeps popping up in the video workflows shared on Reddit and Hugging Face: it's a community fine-tune of MiniMax H3, downloaded hundreds of thousands of times, and on October 21 the V2 version arrived. Before rushing to download 34 GB, though, here's how we see it: with "reworked" video models the temptation is to grab the latest file and hope it runs. It often ends with hours of downloading, memory struggling and, sometimes, a license you hadn't read.

We work with ComfyUI on a perfectly ordinary 12 GB RTX 3060 with 32 GB of RAM, so the first question we ask is always the same: which file makes sense for a card like this, and can I use it for a real job? In this article we lay out what changes in V2, which versions exist (including the 11.8 GB pruned one) and the point almost nobody highlights: the license, which for anyone working from Italy is the real obstacle.

Quick facts

What is MiniMax H3 Singularity and what changes in V2?

Singularity is a "merge" of several official MiniMax H3 checkpoints (the variants for references, for first-last frame and a third intermediate one), then retrained and pruned to reduce flaws. You'll find it on WarmBloodAban's Hugging Face page. A checkpoint is simply the file with the model's "weights": the one you load in ComfyUI.

V2, uploaded on October 21, brings four improvements according to the authors:

• More action and movement: further training on action scenes and motion sequences.
• Better prompt adherence: subject, action, setting and style are respected more; the recommended limit is about 5,000 characters, to be read as a ceiling, not a target.
• Fixed audio reference: in V1, conditioning with an audio clip didn't work well; now the reference audio can also influence the character's color tone.
• More recognizable styles if you name them in the prompt: photorealistic, 3D CG, cinematic CG.

The strengths of the September version 1.3 remain: cleaner movements, sharper faces in medium and long shots, less shiny skin. These are the authors' claims: we found no independent comparisons, so think of it as a promising fine-tune, not a certified leap forward.

Which MiniMax H3 Singularity file should you download with 12 GB of VRAM?

With 12 GB the only sensible file is the 11.8 GB w4a8, because the official V2 currently exists only as a full 34 GB INT8. Fortunately, user WAS has already published the pruned versions of V2. A word on the terms: INT8 means the weights are stored with 8-bit numbers instead of 16 (half the space, little loss); w4a8 brings the weights down to 4 bits, even lighter but less faithful; pruned means part of the model has been simplified to cut weight.

PSNR measures how closely the result resembles that of the full model: the higher, the better. The drop from 29 to 21 dB isn't small, so the w4a8 should be seen as the file "for testing", not for final delivery. Neither repository states how much VRAM is needed: on a 3060 ComfyUI will have to move part of the model to system RAM, and with 32 GB of RAM you're at the limit. If you're short on memory, also look at alternatives designed for that, like FreeVideo, which runs MiniMax H3 with 8 GB of VRAM and 16 GB of RAM. To choose the right card, you'll find our advice in the guide to the best graphics cards for AI.

Can MiniMax H3 Singularity be used for work in Italy?

No, not under the standard license: Singularity is a derivative of MiniMax H3, and the H3 license excludes the European Union. This is the point we care about most. The original card is labeled Apache 2.0, a very permissive license, and that can be misleading. But a fine-tune can't "erase" the license of the model it comes from: in fact, WAS's pruned repository explicitly states the MiniMax H3 Community License.

As Tech Times summarizes, the license in force since August 2 excludes the European Union, the United Kingdom, South Korea and the United States from the licensed territory, even for local use and for generated videos. Outside those countries, companies under 20 million in revenue can use it by displaying the "MiniMax H3" notice; it's also forbidden to use H3 or its outputs to train other models. If you want authorization, you can apply through MiniMax's form. So for an Italian client, the prudent choice for now is another video model: in our handbook of video models for ComfyUI you'll find those with clearer licenses.

How do you use MiniMax H3 Singularity in ComfyUI?

You use it like a regular MiniMax H3 model: put the file in ComfyUI/models/diffusion_models/ and load it with the Load Diffusion Model node, leaving weight_dtype on "default". No custom nodes are needed: H3 support is native. The rest of the workflow (text encoder and VAE, the component that turns the result into frames) is H3's. If it's your first time opening ComfyUI, start with our ComfyUI beginner's guide.

The authors recommend the minimax_h3_ref2v_turbo_4step_v0.1 LoRA (a LoRA is a small file that modifies the model's behavior), which brings generation down to just 4 steps. WAS confirms that "regular" LoRAs also apply to the pruned versions. A good starting point is RunComfy's references→video workflow (designed for v1.3): 8 base steps, 1.5x upscaling and 4 refinement steps. If you want to see H3 on small cards, the video "MiniMax H3 40-Second Video on 6GB VRAM: Can It Really Work?" from the CG Pixel channel shows the trade-offs well.

Try it now

1. First of all, read the license: if the video is for a client or for a monetized channel from Italy, stop here.
2. Check your disk space: even the lightest file weighs 11.8 GB, plus H3's text encoder and VAE.
3. With 12 GB download Minimax-h3_Singularity_ref2va_v2_pruned_w4a8.safetensors; with 16–24 GB consider the w6a8 or the int8 convrot.
4. Add the 4-step turbo LoRA and start with short, low-resolution clips to gauge the times on your card.
5. Write descriptive but not repetitive prompts and name the style ("cinematic CG", "photorealistic"): that's where V2 promises the biggest progress.

Other news

• Mirelo sound effects in ComfyUI: on October 21 the "Video to SFX" and "Text to SFX" partner nodes arrived (PR on GitHub). A paid online service, no downloadable weights.
• VEED nodes: Fabric 1.0, Lipsync 2.0, video background removal and audio cleanup, again as paid partner nodes (PR on GitHub).
• Krea 3 Image "ultra": a new reasoning mode that requires an input image and works at 2K (PR on GitHub).

Frequently asked questions

Is MiniMax H3 Singularity an official MiniMax model?

No. It's a community fine-tune created by the AIGC-Singularity team and published on Hugging Face by WarmBloodAban. It starts from the official MiniMax H3 checkpoints, merges them and retrains them. That's why the improvements are claimed by the authors and not verified by MiniMax.

How big is Singularity V2?

The official V2 INT8 version weighs about 34 GB. The pruned versions published by WAS weigh 21 GB (int8 convrot), 16 GB (w6a8) and 11.8 GB (w4a8). On top of the model weights you need MiniMax H3's text encoder and VAE, so plan for plenty of free disk space.

Why does the card say Apache 2.0 if it can't be used in Europe?

The Apache 2.0 label was chosen by the author of the fine-tune, but Singularity derives from MiniMax H3, which has its own license excluding the European Union. A derivative inherits those limits: WAS's pruned repository acknowledges this too. When in doubt, the more restrictive license applies.

Do you need a custom node to use it in ComfyUI?

No. ComfyUI supports MiniMax H3 natively: just load the file with Load Diffusion Model from the diffusion_models folder and use H3's text encoder and VAE. The 4-step turbo LoRA is optional but recommended to cut down times.

You'll find all our articles and free workflows in the tutorial3d.it ComfyUI section, and step-by-step videos on our YouTube channel @tutorial3d.