MiniMax H3 open weights on Hugging Face: what is in the release

Yes, MiniMax H3 has published weights: the MiniMaxAI/MiniMax-H3 page on Hugging Face describes a 33B-parameter dense, single-stream Transformer released under the MiniMax H3 Community License Agreement, with two task-specific checkpoints. The license has an application form for the USA, EU, UK and South Korea, so read it before you plan production use.

Everything below is from MiniMax's model card, its announcement and the ComfyUI docs, read 2026-09-29. For a hosted call instead of your own GPUs, Sume lists the model as minimax-h3 in the Video generation docs.

What is in the release?

The card describes three modules: H3-Context-IR for input preprocessing, H3-Base for 768p generation and H3-Regenerate-2K for higher resolution. It says the 2K step does not use a conventional super-resolution module; the base model regenerates its own low-resolution result in context, reusing the original multimodal context.

What does it take to run?

The card's serving examples use SGLang with --num-gpus 4 and state BF16 precision; it gives no VRAM figure. It lists SGLang, vLLM, Diffusers and ComfyUI as frameworks. The ComfyUI docs describe a separate text encoder and video and audio VAEs plus optional turbo LoRAs that use 4 to 8 steps, and a default of 20 steps.

What were the announced dates?

MiniMax's announcement is dated 2026-07-31 and says MiniMax planned to open the weights in the coming days, subject to applicable laws and regulations. The model card carries no announcement date, so this post gives none for the weights.

When is a hosted API the easier route?

When you do not want to hold four GPUs or manage a license application. On Sume, POST /v1/videos with model: "minimax-h3" returns a job you poll, at native 480p or 768p with 2K and 4K as priced upscales. That is a different product from the weights: check the license terms for your region either way.