MiniMax H3
Released Jul 31, 2026
MiniMax's omni-modal video model: it reads text, images, video and audio as one context and generates video with native stereo audio, up to 2K and 15 seconds. The 768p base model's weights are on Hugging Face.
- Price per second
- $0.080
- Max clip length
- 15 s
- Max resolution
- 2K
- Native audio
- Yes
- AA arena Elo
- 1158
- LMArena Elo
- 1460
- Capabilities
- text to video, image to video
- Licence
- MiniMax H3 Community License Agreement
Strengths & weaknesses
Strengths
- Native 32 kHz stereo audio with dialogue in 11 languages
- 2K output via the hosted H3-Regenerate-2K stage
- Open weights (two checkpoints, FL2VA and Ref2VA) under a community licence
- $0.08/s at 768P, $0.13/s at 2K
Weaknesses
- Open release is partial: the H3-Context-IR prompt stage and the 2K regeneration stage stay hosted, so local runs are 768p
- The licence excludes the EU, UK, South Korea and US, and products above US$20M annual revenue need written authorisation
- 15-second maximum
Launch post dated 2026-07-31 said the weights would follow 'in the coming days'. The Hugging Face repo MiniMaxAI/MiniMax-H3 now holds the H3-Base FL2VA (text/first-last-frame) and Ref2VA (reference) checkpoints, licence 'minimax-h3-community-license-agreement', matching the LMArena licence tag. The model card says H3-Context-IR is not open-sourced, H3-Regenerate-2K is 'not yet open-sourced', and sparse attention is to come. The licence grants rights only in the 'Applicable Territory', which is worldwide minus the EU, UK, Republic of Korea and USA. Commercial use is allowed there, with 'Powered by MiniMax H3' and 'MiniMax H3' display requirements and prior authorisation above US$20M annual revenue. API audio is an input (reference_audio); output audio is native per the launch post and model card. H3-Omni-Transformer is 33B dense (model card).
Links
API model strings
- minimax
MiniMax-H3