MiniMax ships H3 open-weights video model with day-zero ComfyUI support
MiniMax released H3, an open-weights omni-modal video model, and Comfy, the team behind ComfyUI, built native support for it and shipped it the same day, what the company calls day zero. H3 is MiniMax's third generation of video model, following Hailuo 01 and Hailuo 02, and the first the company has released with open weights.
H3 takes text, images, video, or audio as input and generates video with real stereo sound, up to 2K resolution, and clips up to 15 seconds long. It supports text-to-video from a prompt alone, image-to-video that animates a still picture, first-and-last-frame control where the model fills in the motion between a specified opening and closing shot, and reference-to-video, where supplied images, video, or audio carry a subject, a motion, or a voice through the generated clip. Comfy highlights the model's multimodal context understanding as its lead capability: it can take images, audio, and video together and resolve them against a prompt describing how they relate, collapsing what would otherwise be several separate generation tasks into one model. Motion transfer, where a reference video supplies camera movement or a performance while the subject and style come from elsewhere, is presented as the feature most useful for iterative editing work. Audio is generated together with the video in the same pass, in native stereo, rather than added afterward.
Running a 2K video model locally required substantial optimization work. Comfy found that the model's modulation weights, about 40% of its total parameters, could be pruned and replaced with a functionally equivalent lookup table, shrinking memory use with no loss in output quality by the company's account. The weights also ship with int8 quantization, and custom kernels cut peak VRAM use during inference. Combined with dynamic VRAM offloading, Comfy says these changes cut the model's total memory footprint by 66%, from 123.6 GB at full precision to 42.5 GB for the smallest variants, enough to run on a consumer GPU such as an RTX 3060.
To use it, Comfy instructs users to update ComfyUI to version 0.30.0 or later (or use Comfy Cloud), download the provided workflows or find them in the template library, follow the workflow's instructions to save the model files in the correct directory, then write a prompt, connect any frame or reference inputs, and run. The model weights are published on Hugging Face under Comfy-Org/MiniMax-H3.
Key facts
- MiniMax H3 is an open-weights omni-modal video model, MiniMax's third generation after Hailuo 01 and Hailuo 02, and its first release with open weights.
- It generates video up to 2K resolution and up to 15 seconds long, with native stereo audio produced in the same generation pass, from text, image, video, or audio inputs.
- Comfy shipped native ComfyUI support for H3 the same day it released, calling it day zero.
- Pruning the model's modulation weights (about 40% of total parameters) into a lookup table plus int8 quantization and custom kernels cut total memory footprint 66%, from 123.6 GB to 42.5 GB for the smallest variants.
- Comfy says the reduced footprint lets H3 run locally on a consumer GPU like the RTX 3060; using it requires ComfyUI 0.30.0 or later, or Comfy Cloud.
Why it matters
Open-weights video models with native audio and multimodal input handling are still rare, and MiniMax breaking from its closed Hailuo line to release H3 openly extends that small field. Just as notable is the same-day ComfyUI integration: Comfy did the engineering to make a 2K video model runnable on consumer hardware before most users had even downloaded the weights, which is what turns a research release into something people can actually use the day it lands.
Who it affects
Anyone doing generative video work with ComfyUI gets a new open model to build workflows around, including hobbyists and small studios who lack access to high-end datacenter GPUs. Motion-transfer and reference-to-video features matter most for people iterating on shots rather than generating one-off clips, since a reference video's camera move or performance can be reapplied while the subject and style come from elsewhere.
How to use it
Update ComfyUI to version 0.30.0 or later, or use Comfy Cloud. Download the provided workflows, or find them in ComfyUI's template library, follow the workflow's instructions to place the model files in the correct directory, then write a prompt, connect any frame or reference inputs, and run. The model weights are published on Hugging Face at Comfy-Org/MiniMax-H3. No pricing or licensing terms for the weights are given beyond their being open.
How solid is it
The technical claims, the 2K and 15-second limits, the 40% modulation-weight share, and the 66% memory reduction from 123.6 GB to 42.5 GB, come directly from Comfy's own engineering writeup rather than an independent benchmark. No third-party evaluation of output quality is cited, and Comfy's claim of "no loss in output quality" after pruning is the company's own assessment, not a measured comparison.
Risks and caveats
All performance and quality figures originate with Comfy, the party that built the optimization and has an interest in it working well; there is no outside verification here. The source does not name the individual engineers behind the optimization, attributing it only to "we," and gives no calendar release date, only relative markers like "today" and "this morning." No comparison to other open-weight video models is offered, so how H3 stacks up against alternatives is untested by this account.
“Audio is generated with the video in the same pass, in stereo, not bolted on afterward.”
— Comfy, ComfyUI blog