MiniMax H3: ComfyUI Tool Ecosystem and Up to 27.7× Generation Speedup

Within two weeks of the open weights for MiniMax H3, ComfyUI nodes for long videos, clip stitching, and upscaling to 2560×1440 have emerged, while NVIDIA and Hao AI Lab have demonstrated generation speedups via Sol-Engine (up to 27.7× on GB200) and an open, distilled 4-step FastH3 Preview v1.

Compiled by Sergey KostenchukPublished 2026-09-01Updated 2026-09-18

2026-09-01 Coding NVIDIA

🛠 MiniMax H3: The ComfyUI Tool Ecosystem Is Growing

Within two weeks, ComfyUI nodes for long videos have appeared: Endless-Sampler samples video in chunks with live preview, Extender 2.0 stitches clips together, and UpscaleGuider scales up to 2560×1440. Hao AI Lab has released an open 4-step FastH3 Preview v1.

🌍 The open weights have turned the closed model into a post-training platform: distillation down to 4 steps and a “draft + LTX-refine” scheme yield, according to the authors, up to 27.7× speedups. At MiniMax’s API price of $0.08 per second of 768p, this directly reduces the cost per second of video.

👤 Anyone making videos in ComfyUI gets ready-made nodes for long clips and upscaling to 1440p. Note: FastH3 requires the FastVideo VSA backend, some older Extender 2.0 workflows will need to be rebuilt, and there is no independent verification of the measurements yet.

Source 1: https://haoailab.com/blogs/fasth3-preview Source 2: https://nvlabs.github.io/Sana/Sol-Engine/H3-Super-Acceleration

Sources