MiniMax, a Chinese AI company, released H3 on August 3, 2026. H3 is an AI model that creates short videos — up to 15 seconds long and up to 2K resolution — and it can also create stereo sound to go with the video in the same step. You can give it text, an image, a video clip, or an audio clip, and it will generate video based on that input. H3 is the first video model from MiniMax released as open weights, meaning anyone can download the model files and run the model themselves rather than only accessing it through a paid online service. It follows two earlier MiniMax video models called Hailuo 01 and Hailuo 02 (ComfyUI blog).

MiniMax officially launched H3 on July 31, calling it a general-purpose model that can understand and work with multiple types of input at once — text, video, speech, and music (MiniMax blog). The company's homepage characterizes H3 as open-weight and general-purpose.

Think of an AI model as a very large set of learned patterns stored as numbers. Weights are those numbers. Open weights means the company has published those numbers for anyone to download and use.

The model supports several ways to create video: typing a text description, providing an image, providing a first and last frame (the model fills in everything in between), or providing a reference clip whose style or content guides the output. ComfyUI, a free tool that lets people connect different AI models together in a visual workflow, announced support for H3 on the same day it was released, in version 0.30.0. ComfyUI provides ready-made workflows for three of these methods.

Most AI video tools today create the video first, then add sound as a separate step. H3 creates the audio alongside the video in a single pass. That means the sound is naturally synchronized to what is happening on screen, without needing a second AI model or a manual alignment step.

The main challenge with running a large AI video model on a home computer is memory. H3 in its full form requires about 124 GB of video memory — far more than any consumer graphics card has. ComfyUI's engineering team found two ways to shrink that requirement. First, they identified that about 40% of H3's internal values are modulation weights, a type of parameter that fine-tunes how the model responds to different inputs. They replaced those with a lookup table — a smaller, precomputed reference that produces the same results without storing all the original values. Second, they reduced the precision of some of the model's internal calculations from higher-detail numbers to 8-bit integers, which take up less memory, and wrote custom code to manage memory more efficiently during use (ComfyUI blog).

Together, these changes cut H3's memory requirement by 66%, from 123.6 GB down to 42.5 GB for the smallest version. With a technique called dynamic VRAM offloading, which moves parts of the model in and out of the graphics card's memory as needed, ComfyUI says H3 can run on a GPU as modest as an RTX 3060 — a mid-range graphics card that many people already own. That is a notable threshold: it means you do not need expensive data-center hardware to run a model that can generate 2K video with synchronized stereo sound.

The model files are available in two places on Hugging Face, a popular website for sharing AI models. MiniMax hosts the original version at MiniMaxAI/MiniMax-H3. ComfyUI hosts a repackaged version at Comfy-Org/MiniMax-H3, which includes the memory optimizations described above.

The broader picture is that generative video has been getting better quickly, but most of the best models have only been available through paid online services or required expensive hardware to run on your own. H3's open release, combined with ComfyUI's memory optimizations, changes that. A person with a single consumer graphics card can now download the model, run it on their own computer, and connect it to other AI tools — all without paying for an online service.

The fact that H3 creates sound at the same time as video also matters for people building projects with multiple AI tools. When audio and video come from the same model in one step, there is no risk of the sound being out of sync with the picture. That removes a common source of problems for anyone combining different AI models in a workflow.

ComfyUI's same-day support for H3 is itself worth noting. When a new AI model works with ComfyUI on its release day, it immediately reaches a large community of people who build and share creative projects. Ready-made workflows for three of H3's video modes make it easier for newcomers to get started. MiniMax's choice to release H3 as open weights, rather than keeping it behind a paid service, fits a broader trend over the past year: open AI models are increasingly matching the quality of paid ones, not just offering free access.

For anyone deciding whether to try H3, the comparison is no longer just about which online service produces the best video for the lowest price. It now includes whether you can run the model on your own computer, how easily it connects to other tools, and how much you can modify it. With H3 and ComfyUI's optimizations, that evaluation is practical on hardware many people already have at home.