Video catalogue · Verified local path

MiniMax H3 local guide

Open-weight omni-modal video model with native stereo audio and local ComfyUI workflows for text, keyframes and multimodal references.

Choose an app

Start with Recommended. No terminal commands are shown.

Compare video models
Open ComfyUI setupRecommended verified starting pointOpen Diffusers setupAdvanced Python runtimeOpen official sourcePublisher repository and primary evidence

Official example

Play the official source preview directly below. The media stays on the official publisher server and is never hosted by LocalClaw.

Streamed from the official sourceMiniMax H3 official repository

What it does

Text To VideoImage To VideoReference To VideoAudio Conditioned Video

The official ComfyUI path uses an approximately 21 GB INT8 transformer, a 16 GB NVFP4 text encoder and CPU offloading. A 24 GB NVIDIA GPU with 64 GB system RAM is a conservative practical floor for local 768p previews; the official server example uses four GPUs.

Hardware figures are practical entry floors, not performance guarantees. Resolution, duration, precision, offloading and runtime versions can materially change memory use.

Strengths

  • Native synchronized stereo audio
  • Text, keyframe and multimodal reference workflows
  • Official ComfyUI templates and quantized weights

Limits to know

  • Very large downloads and slow offloaded inference
  • The open local base targets 768p; official 2K regeneration is not yet open-sourced