MiniMax H3: a video model that follows your reference pictures, and where to run it
Community license, weights free to download, commercial use allowed below USD 20 million in yearly revenue. Self-hosting is a big job: a 33-billion-parameter transformer plus a 32-billion-parameter vision-language encoder, about 144GB of weights in BF16, and MiniMax's own example spreads the work across four GPUs. The default license also leaves out the EU, the UK, Korea and the US. In your browser it runs on cnaps studio, included in the basic plan, under a regional license MiniMax granted for our deployment.
It makes a clip and its soundtrack in one pass, and it takes pictures, video and sound as references. Name a reference in the prompt, such as <Picture 1>, and the result keeps that person, product or place; a reference video lends its motion and its audio. Picture and 32 kHz stereo sound come out of the same model, so there is no separate audio pass to line up.
The cost is time. At the default 30 steps a clip takes about 23 minutes in the studio, and clips run from about 5 to 14 seconds. Eight video generation models sit in the cnaps studio catalog, and this is the one for keeping a specific subject consistent from a reference.
What it actually is
An omni-reference audio and video model from MiniMax. One single-stream transformer denoises a packed sequence that holds the text, the reference media and the target video and audio together, which is why picture and sound arrive in sync. Of its 33 billion parameters, about 13 billion sit in branches that can be precomputed and cached. Prompts work reliably in 11 languages.
The open checkpoint is the base model of a three-stage system. MiniMax keeps the prompt rewriting stage and a 2K regeneration stage behind its own service, so output tops out at the base model's resolution.
In cnaps studio you give it a text prompt plus an image, a video or both, and at least one reference is required. The video's soundtrack counts as the audio reference and is cut to the clip length. Length runs from 124 to 345 frames at 24 fps, default 124, and six canvases are offered, default 1344 by 768. Steps run from 20 to 75, default 30. The model is guidance-distilled, so there is no negative prompt.
Try MiniMax H3 in cnaps studio →
The published numbers
MiniMax has published no benchmark scores for H3, and there is no paper yet. What exists is the studio's own timing on the same prompt at different step counts.
Two things follow. Steps change the result, not only the detail, so settle the step count before iterating on the prompt. And about 6 minutes of every run is fixed overhead from streaming the weights on and off the GPU, with roughly 35 seconds per step on top, so fewer steps save less than expected.
Demand is high: 3.6 million downloads last month on Hugging Face, ahead of the 2.1 million for LTX-2.5.
How it compares to the other video generators here
None of these four share a published test, so the table maps inputs and limits. The deciding questions are whether you have a reference to match and how long the clip must be.
Reach for MiniMax H3 when a specific face, product or place has to stay recognisable, or a clip must run past 10 seconds. Reach for LTX-2.5 for 1080p with sound from text alone, Wan2.2 TI2V 5B for quick drafts, and Cosmos3 Nano to continue a clip you already have.
What to chain it with
Text and references in, a clip with its soundtrack out, so it usually sits in the middle of a chain.
Two chains suit it. Before it, Z-Image-Turbo makes the reference picture when you have no photo, which is how the example above was built. After it, SeedVR2 7B can raise the 768-pixel output toward 4K, since the 2K stage is not part of the open release.
Open MiniMax H3 in cnaps studio, start on the 768 by 768 canvas with 124 frames to keep the wait short, and scale up once the reference holds.
FAQ
What does MiniMax H3 need to run locally?
A multi-GPU server. The weights come to about 144GB in BF16, MiniMax's example runs across four GPUs, and no memory figure is published. The default license also excludes the EU, the UK, Korea and the US. In cnaps studio it runs in the browser, included in the basic plan, under MiniMax's regional license.
How do I reference my own image or video?
Connect the image, the video or both, then name them in the prompt as <Picture 1> and <Video 1>; the video's sound becomes <Audio 1>. A reference video should run 2 to 15 seconds and is cut to the clip length, so match it to the frame count you pick.
Can I use MiniMax H3 output commercially?
Yes, within the community license. Commercial use is allowed below USD 20 million in yearly revenue, commercial products must show the name MiniMax H3 prominently, and outputs may not be used to improve other AI models. Larger companies need written authorisation from MiniMax.
