Audio-video AI Β· Wirkung preview

Origami-H3: Big-Screen AI Video with Sound, Folded to Fit Your Laptop

Ask AI β–Ύ

MiniMax H3 can make a video and its soundtrack in one go, but the full model weighs about 45GB and normally wants a large graphics card. Origami-H3 is a preview build from the startup Wirkung that folds the same model down to roughly 10GB and promises to run it straight from your SSD on an ordinary CPU, an Apple Silicon Mac or a laptop chip. It also bakes in a four-step sampler for quick drafts and claims long, unbroken shots. Here is what that could mean for creators, how to try it, what is still unproven, and how to get similar clips online today.

  • About 10GB download
  • CPU and Mac friendly
  • Four-step drafts
Make a video with sound online β†’
Laptop on a night desk playing a neon alley AI video next to a paper origami crane

Uncensored Alternative Video Generator

Image to video

Upload a starting photo, describe the movement, the camera and the sounds you want, and get a short clip in your browser. Adult (18+) fictional characters only.

A starting image is required for video.

The Same H3 Magic, Folded from 45GB to 10GB

Think of the original MiniMax H3 as a huge sheet of paper: powerful, but too big for most desks. Wirkung's idea is to fold that sheet without cutting anything away. According to the model card published on Hugging Face on 11 November 2026, the 21.6-billion-parameter model shrinks from 44.89GB to 10.38GB, about 4.33 times smaller, with no retraining or fine-tuning at all. It keeps the parts that make H3 special: the joint picture-and-sound engine, the speed-up adapter and the decoders that turn numbers back into frames and audio.

For a creator, the promise is simple. You would not need a top-end graphics card to play with one of the most talked-about video-with-sound models. The card says it uses zero graphics memory and streams its weights from the SSD, so it can work on a regular processor, a laptop AI chip or Apple's Metal graphics. Keep one thing in mind: the 4.33 figure is about file size, not speed. Wirkung has not published how long a clip takes to render.

Running MiniMax H3 Without a GPU: CPU, Mac and Low-VRAM PCs

This is the reason most people look this build up. The standard version of H3 is a heavy download that expects a powerful card, which rules out MacBooks, office laptops and gaming PCs with 8GB cards. This folded build changes the target hardware: Linux computers with a modern Intel or AMD processor and Apple Silicon Macs are the two supported platforms, and the 10GB download fits easily on a normal SSD. Windows is not supported yet.

What nobody can tell you yet is how fast it feels. Running a video model on a processor is usually far slower than on a graphics card, and the card gives no timings, so plan for patience rather than instant results. If your goal is low-VRAM H3 right now, the safest picture is this: it is the most hardware-friendly attempt so far, but it is an experiment, not a proven shortcut.

  • Supported: Linux on x86-64 processors and macOS on Apple Silicon (M-series).
  • Download size: about 10.4GB instead of almost 45GB.
  • Needs Python 3.10 or newer; no Windows package at the moment.

Origami-H3 vs MiniMax H3 and Other Faster H3 Builds

The H3 family has grown quickly, and every spin-off takes a different shortcut. VDN-H3 changes how frames look at each other to save time, TaoMate-H3 goes for near real-time streams in three steps, H3 Remix blends weights, and Pika H3 Flex adds trained adapters. This one takes another road entirely: make the model smaller and move it onto everyday hardware. The table compares it with the original H3 using only the numbers each side has published.

Original MiniMax H3Folded build (preview)
Download sizeAbout 44.89GBAbout 10.38GB
Graphics memoryA large graphics cardClaimed 0GB, streams from SSD
Where it runsMainly graphics cardsCPU, Apple Metal, laptop AI chips
Sampling stepsStandard multi-step, Turbo LoRA optionalFour steps with Turbo LoRA built in
Long clipsUsually made in segmentsClaimed constant memory for unbroken takes
Tool supportComfyUI and other common toolsOnly the Wirkung runtime

Long Uncut Takes and Four-Step Drafts: Fun Ways to Use It

Two claims make this build exciting for storytellers. First, four-step sampling: instead of waiting for dozens of passes, you can sketch several versions of a scene and pick the one you like. Second, what Wirkung calls an O(1) horizon carrier, which in plain words means the model keeps a fixed-size memory while it goes on, so a long shot with continuous narration should not need to be cut into five-second pieces and glued back together.

No public sample videos exist yet, so treat these as ideas to test rather than guaranteed results. Because it all runs on your own machine, your footage and prompts never leave your computer, which is a real plus for private projects and unreleased ad concepts.

  • One-take city walks: a camera gliding through rainy neon streets with ambient sound that never jumps.
  • Storyboard drafts on a MacBook: rough versions of each shot with sound effects before a big render.
  • Short drama and ad tests: four quick takes of the same line, then refine the best one.
  • Live-style scenes: a stream-like view, such as a beach at sunset or a crowded market, that keeps rolling.
Continuous cinematic camera glide through a rainy neon city street at night

Download and First Run in Three Commands

Getting started is short. Install Wirkung's Python package, pull the model with its own command-line tool, then check that the files arrived intact. The model card's sample code stops at printing model information, and Wirkung has not yet documented the command that actually renders a video, so expect to experiment. The weights ship as 307 special .o3f files that only the Wirkung runtime can open, so ComfyUI and Diffusers workflows do not work with it.

1. Install the runtime
pip install -U wirkung
2. Download the model
wkng pull wirkung/Origami-H3 -o ./model
3. Check the files
wkng verify ./model
wkng info ./model

Is Origami-H3 Legit? Claims vs What Is Proven

Zero graphics memory, four steps and endless takes sound almost too good, so it is fair to ask whether it is the real deal. The verifiable part checks out: the Hugging Face repository exists, its storage is about 10.39GB, and the package appeared on PyPI in early November 2026. The rest is the developer's own description. There are no benchmark scores, no sample clips and no independent tests yet; quality numbers are pointed to a paper on the repository.

Community reaction so far is cautious. In the only discussion thread on the model page, a user called the claims amazing but said they did not want to install a brand-new pip package without source code, and the question had no answer at the time of writing. The runtime is closed, so a careful approach is to test on a spare machine and wait for others to share results.

What Wirkung saysIndependently confirmed?
Size around 10GB10.38GB from 44.89GBYes, repository storage matches
No graphics memory needed0GB, weights read from SSDNot yet
Same quality as H3No retraining neededNot yet, no scores or samples
Long unbroken clipsConstant memory over timeNot yet, no sample videos
Open source runtimeNot offeredClosed, as users have noted

Prompt Templates for H3-Style Video with Sound

Wirkung has not published a prompt guide, so these follow the general way people prompt H3-family models: name the subject, the action, the camera and the sound. For four-step drafts, keep sentences short and skip piles of adjectives so the result stays stable.

Long take
One continuous uncut shot, camera slowly dollies through a rain-soaked neon alley at night, puddle reflections, distant thunder and soft synth hum in the audio, no scene cuts.
Short drama line
Medium close-up, a woman in a red trench coat turns to camera and says "You were never supposed to find this." Tense string music, ambient city noise, cinematic lighting.
Product ad
Macro shot of a perfume bottle rotating on black glass, light sweeps across, crisp glass clink sound, elegant piano sting.
Beach walk (18+)
Adult fashion model (25+, fictional) in a white bikini walking along a sunset beach, slow tracking shot, waves and gentle breeze audio, golden hour.
Runway look (18+)
Fictional adult model in a lingerie runway look, catwalk lights, camera pans up, bass-heavy runway music.

Uncensored Origami-H3: Adult Video Ideas Without the Wait

There is no uncensored or NSFW version of the folded model, and its rules forbid modifying it, so community variants are unlikely to appear soon. Our image-to-video tool offers similar freedom for adult (18+) clips online, with no 10GB download and no setup.

Start from a photo of a fictional adult character, describe the scene and the sound, and you can try several takes in minutes. The prompt templates above work well as a starting point, and keeping one reference picture across shots helps a character stay consistent from scene to scene.

Bikini beach walks

Golden-hour strolls along the waves for adult models, with breeze and surf sounds.

Lingerie runway looks

Catwalk lights, a slow camera pan and heavy bass for fictional adult characters.

Nightclub mood dances

Neon haze, a glance over the shoulder and a pounding beat.

  1. Use a sharp, well-lit starting image of a fictional adult (18+) character.
  2. Describe what should change: motion, camera move and sound, not the outfit again.
  3. Generate a few short takes and keep the best one, the same idea as four-step drafts.

Turn a still photo into a moving clip with sound, right in your browser.

Try uncensored image to video
Adult model in a white bikini walking along a sunset beach for an AI video scene

Self-hosting vs Generating Online on Spicy AI

Running it yourself keeps everything on your computer; the online route skips the download and setup.

Local deployment (official weights)Spicy AI online
Content filtersDepends on your own setupRelaxed filters for adult content (18+)
Open weightsDownloadable, about 10GBNot needed, runs in the browser
Model in useThe folded H3 previewA similar hosted video model
Getting startedLinux or Apple Silicon Mac, Python and a downloadOpen the page and upload a picture
  • Image-to-video from your own starting photo
  • Prompted motion, camera moves and sound
  • Results saved to your history

FAQ

What is Origami-H3?

It is a preview build of MiniMax H3 released by Wirkung on Hugging Face on 11 November 2026. It keeps the 21.6-billion-parameter audio-video model but shrinks the files from 44.89GB to 10.38GB without retraining.

Does it really run without a graphics card?

Wirkung says it uses 0GB of graphics memory and runs from your SSD on a CPU, a laptop AI chip or Apple Metal. No render times or quality results have been published, and nobody has independently confirmed it yet.

Is 4.33x smaller the same as 4.33x faster?

No. The 4.33 figure compares file sizes. Speed has not been published.

Can I use it on Windows or in ComfyUI?

Not right now. The package is only built for Linux on x86-64 processors and for Apple Silicon Macs, and the weights use a special format that only the Wirkung runtime opens.

Is the picture as good as the original H3?

Unknown. The model card shares no quality scores or sample clips, so compare results yourself before relying on it.

How do I get H3-style video with sound without installing anything?

Use an online image-to-video generator. Upload a starting picture, describe the motion and sound, and the clip renders in your browser.

Sources

All uncensored AI models β†’

Sign in

Welcome to Spicy AI

or