AI Video Generator
Imagifly — MiniMax H3
MiniMax H3 is a multimodal generation model MiniMax released on 31 July 2026, and it is not part of the Hailuo line: Hailuo 01 through Hailuo 2.3 are the earlier video products, while H3 is a new general text/image-to-audio-video model and the first MiniMax video model released with open weights.
- Image
- Video
- Audio
- MiniMax
From 22 credits per video

About MiniMax H3
MiniMax H3 is a multimodal generation model MiniMax released on 31 July 2026, and it is not part of the Hailuo line: Hailuo 01 through Hailuo 2.3 are the earlier video products, while H3 is a new general text/image-to-audio-video model and the first MiniMax video model released with open weights.
It is a 33B dense single-stream Omni-Transformer with a Qwen3-VL-32B text encoder. The published checkpoints are CFG-distilled and split into H3-Base-FL2VA and H3-Base-Ref2VA. It handles text-to-audio-video, first/last-frame control and reference-driven generation from images, video or audio, with native 32kHz stereo, 4-15 second clips and 768p by default; 2K output requires the API-side H3-Regenerate-2K module.
Weights are published in BF16 on Hugging Face with diffusers, SGLang, vLLM and ComfyUI support, but under the MiniMax H3 Community License dated 2 August 2026 rather than a standard open-source licence. The licence excludes the European Union, the United Kingdom, South Korea and the United States, where a separate application is required. H3-Context-IR and the 2K regeneration module stay API-only.
What it is good at
- Native audio and video generated together, 32kHz stereo
- 4-15 seconds per generation, 24 FPS, 768p by default, 2K via an API module
- 33B dense single-stream Omni-Transformer with published BF16 weights
- Community licence excludes the EU, UK, South Korea and the USA
Technical specs
- Developer
- MiniMax
- Released
- 2026-07-31
- Latest version
- MiniMax-H3 (API: H3 / Highspeed / Lite / Max)
- License
- MiniMax H3 Community License (excludes EU, UK, South Korea, USA)
- Architecture
- 33B dense single-stream Omni-Transformer with a Qwen3-VL-32B text encoder
- Parameters
- 33B (BF16 weights published)
- Task
- Text-to-audio-video, first/last frame and reference-driven video
- Recommended settings
- API variants: MiniMax-H3 (768P/2K, 4-15s), MiniMax-H3-Highspeed, MiniMax-H3-Lite and MiniMax-H3-Max (5-15s). Reference limits are roughly 9 images, 3 videos totalling 15s, 3 audio clips and 12 files overall. No sampler step count is published.