What it does
MiniMax H3 is a general-purpose generative system that understands text, images, video and audio together and produces video with native stereo sound. Clips run 4 to 15 seconds at 24 fps in many aspect ratios, with a default 768-pixel short side and 2K available through a separate regeneration step. Dialogue is stably supported in 11 languages. The open release has two checkpoints. FL2VA covers text-to-video and generation from a first frame, a last frame or both. Ref2VA takes reference inputs, up to 9 images, 3 video clips and 3 audio clips, with at most 12 files in total. The full weights are released, including for fine-tuning, and the repository's SGLang example starts with four GPUs. Not everything is in the open release. The context-processing stage (H3-Context-IR) and the 2K regeneration stage (H3-Regenerate-2K) are reached through the MiniMax Open Platform API, so a fully local setup produces 768p output. The sparse-attention code is also not yet published. The repository bundles a prompt-writing skill for agent tools plus eight video skills built for MiniMax Hub's canvas workflow. H3 is also offered through an online API, a web app and a desktop app. The repository refers to a MiniMax H3 Community License that governs use, and submitted content is moderated automatically.
Capabilities
- Video with native stereo audio, 4 to 15 seconds at 24 fps
- Text-to-video, first and last frame, and multi-reference modes
- Full weights released as two task-specific checkpoints
- Local serving through SGLang, vLLM, diffusers and ComfyUI
- Bundled prompt-writing skill for agent tools
When to use it
- You want to run an open video-and-audio model on your own GPUs
- You need to steer a video with reference images, clips or audio
- You plan to fine-tune a video model starting from released weights
- You would rather use the hosted API or app than deploy anything yourself
