Who makes MiniMax-H3-TAE?

MiniMax-H3-TAE is made by Kijai.

Does MiniMax-H3-TAE work in the EU?

Available in the EU. Apache-2.0 weights are freely downloadable from Hugging Face with no regional restriction in the license; runs locally in ComfyUI. Last verified 14 Nov 2026.

How much does MiniMax-H3-TAE cost?

Open source. Free to download from Hugging Face under Apache-2.0. Licence: Apache-2.0.

Where was MiniMax-H3-TAE shown?

In 1 creator video by AI Search. Most recently by AI Search on 12 Aug 2026, at 2:24 in "The BEST local AI video generator just got BETTER!".

MiniMax-H3-TAE

by Kijai

Try it Hugging Face

Small 2D tiny VAE for live latent previews of MiniMax-H3 video generation in ComfyUI via KJNodes.

Available in EUPlatforms & Infra Video GenerationModel

Updated 14 Nov 2026EU status checked 14 Nov 2026Last shown 12 Aug 2026

MiniMax-H3-TAE at a glance

What is MiniMax-H3-TAE?: MiniMax-H3-TAE is a model from Kijai. Small 2D tiny VAE for live latent previews of MiniMax-H3 video generation in ComfyUI via KJNodes.
Who makes MiniMax-H3-TAE?: MiniMax-H3-TAE is made by Kijai.
Does MiniMax-H3-TAE work in the EU?: Available in the EU. Apache-2.0 weights are freely downloadable from Hugging Face with no regional restriction in the license; runs locally in ComfyUI. Last verified 14 Nov 2026.
How much does MiniMax-H3-TAE cost?: Open source. Free to download from Hugging Face under Apache-2.0. Licence: Apache-2.0.
Where was MiniMax-H3-TAE shown?: In 1 creator video by AI Search. Most recently by AI Search on 12 Aug 2026, at 2:24 in “The BEST local AI video generator just got BETTER!”.

Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.

About

MiniMax-H3-TAE is a tiny autoencoder (TAE) checkpoint published by Kijai on Hugging Face. It is a quickly trained 2D tiny VAE for MiniMax-H3, intended to decode latents into rough RGB frames so that video generation can be previewed while sampling runs. The model card says the result is not the best, but it still beats the default latent2rgb preview.

It currently works only with the ModelPreviewOverride node in ComfyUI-KJNodes. A YouTube creator (AI Search) describes the file as under 10 MB and says it lets users watch the generation live and cancel early if the output looks poor.

The card also notes that madebyollin has trained a more proper version (taeh3.safetensors, in the taehv repository), which also works with Kijai's node. This is a preview helper, not a generation model.

Key features