MiniMax H3
Video modelOpen weightsCard summarizedAn Open Model Breaking the Boundaries Between Tasks and Modalities
● Re-checked 12 Oct 2026 · 68 facts · 2 corrections · 3 unconfirmedSource: MiniMax-H3 model card (Hugging Face), with the H3 launch post and open-source announcement
Where to call it
- MiniMax API
MiniMax-H3 - Self-host (weights)
MiniMaxAI/MiniMax-H3
Good for
- Instruction following, accurate text and brand rendering, and video-to-video motion transfer
- Reference-driven generation from up to 9 images, 3 video clips and 3 audio clips (12 files)
- Voice, sound effects and music generated jointly with the picture
Overview
H3
An Open Model Breaking the Boundaries Between Tasks and Modalities
- Max resolution
- 2K
- Max length
- 15 s
- Modes
- text-to-video, image-to-video
- Released
- 31 Jul 2026
MiniMax H3 is a general-purpose omni-modal generation model: it reads text, image, video and audio context and generates video with native stereo audio, up to 15 s at 2K. The open release is a 33B-parameter dense single-stream transformer in two task checkpoints that outputs 768p; 2K comes from a separate regeneration module and prompt enhancement from H3-Context-IR, both API-only. MiniMax publishes no benchmark numbers; a technical report is promised. model card; launch post
MiniMax-H3- Typical price
- $0.08 / sec
- Max resolution
- 2K
- Max length
- 15 s
- Aspect ratios
- 21:9, 16:9, 4:3, 1:1, 3:4, 9:16
- Frame rate
- 24
- Released
- 31 Jul 2026
- text-to-video
- image-to-video
- first and last frame
- reference-to-video (up to 9 images, 3 videos, 3 audio clips)
- video editing
- 2K regeneration