Day 0 support
MiniMax H3 open weights, running locally on your own GPU
H3 generates video and its 32 kHz stereo soundtrack in a single denoising pass. One transformer, picture and sound out together, so a take is a finished clip rather than a silent one waiting for an audio tool.
Four nodes on the canvas, no per render cost, and nothing leaves your machine.
Hand-drawn Japanese TV anime, 1970s cel look: flat bold colour, thick confident ink lines, speed lines and impact frames, warm afternoon light. Rapid hard cuts, no dissolves. [0.0-2.0s] A round blue robot cat with a white face and a golden collar bell YANKS a glowing gadget from the pouch on its belly.
Live-action science fiction, blockbuster trailer grade: anamorphic lens, deep teal and ember-orange, volumetric haze, heavy 35mm grain, handheld urgency. Hard cuts only. [0.0-2.0s] Klaxons. A starship corridor strobes red. An engineer in a scuffed pressure suit SPRINTS at the camera, sparks bursting from ruptured conduits either side.
Breathtaking FPV drone cinematography, ultra wide lens, hyper-real natural colour, extremely fast forward flight, aggressive banking. No people. Continuous unbroken shot. The camera SCREAMS down a granite spine at high speed just metres above the rock, snow spraying off the ridge in its wash, then rolls hard left and DIVES into a cloud-filled valley.
Japanese prime-time variety television, multi-camera, glossy saturated broadcast lighting, bold on-screen graphics, fast energetic cutting between cameras. [0.0-2.5s] Wide: a bright studio set explodes with confetti cannons. A presenter in a sharp blazer throws both arms up and shouts an excited introduction to camera, the audience roaring.
Photoreal rocket launch documentary, ultra high dynamic range, long-lens and close tracking cameras, night, enormous scale. Hard cuts. [0.0-1.5s] Extreme close on an engine bell in darkness. IGNITION: a blinding white-orange torch erupts directly at the lens, the frame blowing out.
The exact prompt sits under each clip. Every one is 24 fps and carries its own generated soundtrack, muted here for autoplay.
What it measures
Rendered, not projected. Everything below came off a real run.
Where the memory goes
Gigabytes, measured on an L40S. The lighter bar is what the file weighs, the accent bar is what is actually resident once loaded.
Four nodes, because the inputs differ
A node should say what it takes. Wiring order on the reference node is the numbering the prompt addresses, so the order you connect things in is meaning, not decoration.
Text to Video
A prompt and nothing else. One click from an empty canvas to a finished clip.
Image to Video
Bring a still you already like into motion. Wire it into the first frame.
First and Last Frame
Pin the opening frame, the closing frame, or both, and let the model fill the middle.
Reference to Video
Carry a character, a look, a camera move or a voice across shots. Up to nine images, three clips and three pieces of audio, addressed by position in the prompt.
How it fits
Cut 26 GB out of the transformer
66.3 GB → 40.3 GBEvery block carries a 96,768 by 2,688 projection for modulation, and it only ever sees the timestep. One number in means it traces one smooth curve, so we sampled that curve and took its SVD.
- Five real dimensions out of 2,688. Everything after the fifth singular value is numerical dust.
- An eight column basis keeps all of it, so each block drops from 520 MB to 1.5 MB.
- Not lossy. The input has nothing outside that subspace, so the projection discards nothing that was there.
- Derived from the original bf16 weights at load, so no repacked format is involved and a future checkpoint needs no code change.
- No lookup table to index, so any timestep is exact rather than snapped to the nearest stored row. The sampler stays free.
Made every step 5.9 times faster
14.9 s → 2.52 s per stepH3 runs a 32B conditioner beside its denoiser. Left resident it takes 19.5 GB, which forces the denoiser to cross the bus on every step of every render.
- The card sat under a third utilisation. It was waiting, not computing.
- The prompt is encoded once, so the two never actually need the card at the same time.
- The run is split in two and the conditioner steps aside before denoising starts.
- Utilisation went from under a third to 100 percent.
Verified every node by rendering it
10 to 23x closer than an ignored keyframeA keyframe that gets silently dropped still produces a clip that plays perfectly, so watching the output proves nothing. Each node is measured against its own inputs instead.
- The first rendered frame is compared to the image wired into it.
- Ours measure 0.0217 and 0.0092. Two unrelated images measure 0.2157.
- The four input paths do not share failure modes, so one passing tells you nothing about the others.
Running it
- 1
Open Inline Studio and pick the MiniMax H3 card
It is the first card on the getting started screen. Clicking it lays down a prompt node wired into an H3 node, with a prompt and settings already filled in, so there is nothing to configure before the first render.
- 2
Let the node fetch its weights
The node face shows a models missing hint. Click it and the popup downloads what H3 needs and puts each file where the engine looks for it. This is a large download and the popup shows the size of every part before you start.
- 3
Press Run
The take that lands on the canvas is one MP4 with its soundtrack already muxed in. Generating again adds a take rather than overwriting the last one, so nothing you liked is lost.
You can train a LoRA for it
Every H3 node has a LoRA input, and the Trainer tab builds the adapter. Train on still images to learn look, style, character and lighting, or on short clips from about a second up to learn motion as well. Both write the same file, and a single dataset can hold both. Wire the finished adapter into Text to Video, Image to Video, First and Last Frame, or Reference to Video, all four.
The base is frozen at 4-bit while the adapter trains in full precision, and the modulation branch is factorised first, which is what takes the transformer from 66GB to about 40GB before quantisation even starts. Latents and captions are cached in two separate passes, because the video VAE and the 32B conditioner cannot sit on the card at the same time. Measured on an L40S, a run peaks at 20.6GB and the caption pass is what sets it, while training itself sits at 11.7GB. 24GB is comfortable. A 16GB card works too, because the conditioner runs on the CPU there and the peak drops to 12.7GB, but a step goes from 1.8s to 16s and the caption pass takes 19 minutes. Clips cost the same VRAM as stills, about 20.4GB whether the clip is 0.92s or 4.5s, because the caption pass sets the peak either way. Steps are quick, about 0.6s at 512px on an L40S and 0.72s on an RTX PRO 4500 Blackwell, but the 62GB checkpoint takes roughly seven minutes to stream and quantise before the first one.
How to train a LoRA, step by stepQuestions people ask first
What hardware do I need to run MiniMax H3 locally?
System RAM is the binding constraint rather than the GPU. Peak unreclaimable host memory measured 46.7 GB across a full run, so 64 GB is comfortable. On the GPU side, five clips at 960x544 peaked at 38.9 GB of VRAM, identical to three significant figures across all of them. Smaller canvases need considerably less, and the reference node needs more than the other three.
Does MiniMax H3 generate audio as well as video?
Yes. One transformer denoises the picture and a 32 kHz stereo soundtrack in the same pass, so a finished take is a single MP4 with the audio already muxed in. It is not a video model with a vocoder added afterwards.
How long does a clip take to generate?
About 7.2 minutes for a 10 second clip at 960x544 on an L40S, measured across five clips that landed within 7 percent of each other. A denoising step is 2.52 seconds at 608x352. Canvas is by far the biggest lever: time grows faster than pixel count because attention cost is superlinear, so doubling the canvas costs more than double.
Can I run it on a 12 GB card?
It will run with the denoiser streamed from system RAM rather than kept on the card, which works but is much slower. The engine picks that automatically when the card cannot hold the weights, so there is no flag to set.
Do I have to convert or repack the weights?
No. Inline Studio loads the original bf16 checkpoints and derives everything it needs at load time, so there is no dependency on a particular repacked file format and a future checkpoint loads without a code change.
Is there a smaller download than 66.3 GB?
Yes. The pruned and fp8 community builds load as well, and the fp8 one is 21.0 GB for the same model. They ship the modulation branch already reduced, which is the same reduction Inline Studio would otherwise derive at load, so nothing is lost by taking the smaller file. It saves download and disk rather than memory: every build occupies the same amount once loaded. The int8 builds are the exception and are refused, because their weights are stored rotated and that transform cannot be undone outside the tool that applied it. LoRA training needs the bf16 file.
What are the four nodes for?
Text to Video takes a prompt alone. Image to Video animates a still wired into the first frame. First and Last Frame pins either end, or both. Reference to Video carries a character, look or voice across shots using up to nine images, three video clips and three audio clips, addressed by position in the prompt.
Is it free, and does anything leave my machine?
Inline Studio is free and open source under GPL-3.0, and local generation runs entirely on your own hardware with no account and no external server. The H3 weights are licensed separately by MiniMax and carry their own terms, so read those before you rely on them.
What licence do the MiniMax H3 weights come under?
The MiniMax H3 Community License Agreement, which is not a standard open source licence. It excludes four territories, sets a revenue threshold above which you need written authorisation, and requires commercial products to display the model name. The terms are broken out in the licence section on this page, and you should read the licence itself before relying on it.
About the weights licence
Inline Studio ships no model weights. You bring your own. The app is GPL-3.0. H3 is licensed separately by MiniMax, under a community licence rather than a standard open source one. Three points stand out, and none of this is legal advice:
- It applies worldwide except in the European Union, the United Kingdom, the Republic of Korea and the United States of America, which the licence names as Excluded Territories.
- Commercial products over 20 million US dollars in yearly revenue need separate written authorisation from MiniMax.
- Commercial products using it must display MiniMax H3 in their interface, and an acceptable use policy applies.
Be honest with yourself about the hardware
This is a 33B transformer sitting beside a 32B conditioner. The download is large and system RAM is the binding constraint, not the card. Plan on a 24 GB GPU and 64 GB of system RAM. Canvas is the biggest lever on speed, far more than any other setting.
If that is not your machine, the same model is one click away as an API node with nothing to set up. That is a perfectly reasonable thing to use instead.