MiniMax H3 open weights run locally in ComfyUI

MiniMax H3Open weights (MiniMax H3 Community License)

MiniMax H3 is MiniMax's first open-weights video model, making clips with stereo sound up to 2K and 15 seconds. ComfyUI supported it on day one with pruned, quantized files that can run on a 12 GB RTX 3060.

What it does

MiniMax H3 takes text, images, video and audio and makes video with native stereo sound in the same pass. It handles text-to-video, image-to-video, first-and-last-frame, and reference-to-video, including motion transfer from a reference clip.

ComfyUI shipped T2V, I2V and R2V workflow templates on launch day. Comfy pruned the modulation weights and added int8 ConvRot quantization, cutting the memory footprint from 123.6 GB to 42.5 GB, and dynamic VRAM offloading does the rest.

Results and limits

Comfy says the pruning causes no loss in output quality, which is its own claim. In an owner-reported run cited by a Smeltcore recipe, an RTX 3060 12 GB took about 20 minutes for 3 seconds at 1376×768 and 24 steps, or 4.5 minutes for 5 seconds at 864×480 with an 8-step Turbo LoRA.

The recipe says host RAM is the real bottleneck: plan on 32 GB with --disable-pinned-memory (64 GB without it) and about 45 GB of NVMe space.

Worth knowing

Users in the EU, UK, US and South Korea need to request a license from MiniMax via the H3 License Request Form (platform.minimax.io/h3-license) before use. Commercial use of local outputs is licensed separately through Comfy. You need ComfyUI 0.30.0 or newer.

Try it yourself

In ComfyUI 0.30+, open Template Library > Video > MiniMax H3 Image to Video, let it download the models, then prompt: "The scene opens exactly on image 1. [Action], camera [move]. Audio: [ambience], [effect]." Start at 864×480 with the Turbo LoRA.
Video generationVoice and audioWorkflowsLocal
Read the primary source >

More coverage

Related entries

< Back to all entries