The short version: you can run MiniMax H3 on your own hardware if you have two 24GB GPUs, about 200GB of system RAM and 270GB of free disk, and if you live somewhere other than the United States, the European Union, the United Kingdom or South Korea. The weights are open. The word "open" is doing less work than usual.
H3, which MiniMax also calls Hailuo 3.0, launched at the end of July and is the most capable video model anyone has handed out as a download. It renders native 2K at 24fps, runs up to 15 seconds a clip, generates 32kHz stereo audio in the same pass, and takes up to 9 images, 3 videos and 3 audio files as references. The hosted API charges $0.14 a second at 2K and $0.10 at 768p. The weights, published to Hugging Face in early August, promise the same model with no meter running. Two things stand between a filmmaker and that promise: a hardware bill and a license.
What it takes to run MiniMax H3 locally
The vLLM team's deployment recipe is the most honest spec sheet published so far. The model ships as two task-specific partitions: FL2VA for text and first-frame to video, and Ref2VA for reference-driven generation. Each is about 134GB of BF16 safetensors. Both together need 270GB of disk before you have rendered a frame.
The validated consumer setup is two GPUs, either a pair of RTX 4090s at 24GB each or RTX 5090s at 32GB, running tensor parallel with layerwise offload to host memory. Hence the recipe asks for at least 200GB of system RAM and recommends a 384GB host. The four-GPU configuration that serves both partitions at once is not something anyone builds under a desk.
Community GGUF quantizations get the transformer down to 15.6GB at Q3, but the text encoder and VAEs add 20GB or more on top and nobody has benchmarked those builds properly. Treat the "runs on an 8GB card" posts as aspirational.
On September 1, vLLM published numbers for FastH3, a four-step distilled student of H3: a complete 10.1-second MP4 with audio in about 8.7 seconds, faster than playback. That ran on eight B300 GPUs, so it previews hosted H3 next year and says nothing about your workstation tonight.
The license draws a map
The MiniMax H3 Community License covers a territory defined as worldwide excluding the United States, the European Union, the United Kingdom and South Korea. If you are in one of those places, the standard license does not authorize you to deploy the weights. Organizations with $20 million or more in revenue need separate authorization anywhere.
MiniMax's head of developer relations, Ryan Lee, said the US carve-out follows ongoing generative video copyright litigation with major Hollywood studios. That litigation is the suit Disney, Universal and Warner Bros. Discovery filed in September 2025, which accuses Hailuo of generating Darth Vader, Minions and Wonder Woman on request. A federal judge denied MiniMax's motion to dismiss on May 22, 2026, so the case is live and the geography of the license will follow it.
There is a door. MiniMax says users in the excluded regions can apply for individual authorization by email if they commit to a content compliance mechanism that meets local law. For a studio with lawyers that is paperwork. For a solo filmmaker in Lyon or Austin it is a reason to look at the hosted API, where the geography does not apply and the meter does.
What the meter actually costs
Run the arithmetic before buying a second 4090. A 90-second short built from 12 shots of 8 seconds, at the three takes per usable shot most creators report, is 288 seconds of generation. At 2K on the H3 API that is about $40. At 768p, under $29. A two-GPU workstation with 384GB of RAM costs several thousand dollars before it renders anything. Local wins when you are iterating hundreds of takes a week, or when the footage cannot leave the building. It loses on almost everything else.
The clean alternatives
If the license rules you out and you still want weights on disk, two models are worth the download. Wan 2.2 is Apache 2.0, as unrestricted as it gets, and the 5B variant runs from 8GB of VRAM with offloading. LTX-2.3, released by Lightricks on March 5, is the only other open model with native synchronized audio, runs on a 12GB card, and its community license allows commercial use under $10 million in annual revenue. Neither matches H3 on motion, by most side-by-side reports. Both let you ship without an email to Shanghai.
What to do about it
Read the license before the README. Then decide whether you live in a territory it covers, whether you can build the rig, and whether you generate enough footage for the rig to pay for itself. If any answer is no, use H3 through a hosted API as one model in a multi-model workflow, the way we argued after the Sora shutdown. Promvie's pipeline runs on hosted models for exactly this reason. The weights stay someone else's problem, and so does the map.
Open weights used to mean you could run it. With H3 they mean you can run it if you live in the right place and own the right machine, which is a narrower promise than the download button suggests.