Overview
MiniMax H3 is an open omnimodal generation system from MiniMax. It puts understanding and generation inside one model: the input can be a multimodal context mixing text, images, video and audio, and the output is video up to 2K and 15 seconds with native stereo audio.
The spec sheet reads 24 FPS, 32kHz native stereo, stable support for 11 languages, and aspect ratios from 21:9 to 9:16. Two builds split the work. H3-Base-FL2VA covers text to video, first frame, last frame and first plus last frame. H3-Base-Ref2VA takes omni-references of up to nine images, three videos and three audio clips.
What the community did after release deserves its own paragraph. MiniMax collected 11 new integrations, including 8-step reference-conditioned video with audio generation, 4-step distillation, sparse attention and reference token compression, served across GPU servers, TPUs and Apple Silicon. One developer used an agent to gather optimisations and pushed real-time generation onto a single RTX 5090, rendering 30 seconds of video in 23.51 seconds, about 1.28 times real time.
Key Features
- Omnimodal input understanding: Handles multimodal context built from text, images, video and audio, following complex multimodal instructions rather than a single text prompt.
- 2K, 15 seconds, native stereo: Up to 2K resolution and 15 seconds at 24 FPS, with native 32kHz stereo audio rather than a track added afterwards.
- Two builds, two jobs: H3-Base-FL2VA covers text to video, first frame, last frame and both frames. H3-Base-Ref2VA accepts up to nine images, three videos and three audio clips as references for character and scene consistency.
- 11 languages, many aspect ratios: Stable support for Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish and Arabic, with ratios from 21:9 and 16:9 through 4:3, 1:1, 3:4 and 9:16.
- Community pushes real time onto one card: The 11 integrations include 8-step reference generation, 4-step distillation, sparse attention and reference token compression. A community build runs on a single RTX 5090 at 23.51 seconds for 30 seconds of video.
- Open weights plus a hosted API: Weights are open for local deployment and customisation, while teams that would rather not run infrastructure can use MiniMax's per-second commercial API.
Use Cases
- Generate short video clips with sound effects based on text descriptions
- Input first and last frame images to generate transition animation videos
- Provide reference videos and audio to generate multimodal content with consistent style
- Create multilingual dialogue videos for cross-language content creation
- Generate videos with different aspect ratios to suit social media or film production
Pros
- Open source and free, can be deployed locally
- Full-modal input, strong understanding capability
- Output with native stereo audio, strong immersion
- Supports multiple aspect ratios and resolutions, high flexibility
- Multilingual support, suitable for international applications
Pricing
Weights are open and free for local deployment and customisation, with commercial terms following the MiniMax open source license. The official MiniMax API is billed per second, at 0.4 CNY per second for 2K and 0.23 CNY per second for 768P.
Summary
H3 matters because it puts omnimodal understanding and video generation in one model. Most video models take text or a still image; H3 also takes video and audio as references, which helps when a shot has to keep a character or a visual style consistent.
The open release is the other half. Among the 11 community integrations are distillation and sparse attention that push inference cost down, and a single-card real-time build already exists. For budget-sensitive teams that means an alternative to per-second pricing. Deployment is still demanding, since 2K output and long audio have real memory requirements.
Version History
- 11 new community integrations (2026-10-26): MiniMax collected 11 new community integrations since the open release: 8-step reference-conditioned video with audio generation, 4-step distillation, sparse attention and reference token compression, served across GPU servers, TPUs and Apple Silicon with matching workflow integrations.
- Community runs real-time video on a single RTX 5090 (2026-10-08): A ComfyUI blog author used agents to gather optimizations and fit MiniMax H3 real-time video generation onto a single RTX 5090 with 32GB of VRAM, producing 30 seconds of video in 23.51 seconds, about 1.28x real-time throughput.
- Hailuo 3 video model enters commercial use with open weights (2026-09-01): MiniMax's Hailuo 3 video model entered commercial use at 0.4 CNY per second for 2K and 0.23 CNY per second for 768P, the domestic video model of its period to open its weights for local deployment and customization. It passed 24 million downloads and more than 300 derivative videos within three weeks of release. Hailuo 3 Max was also integrated into the MiniMax open platform and MiniMax Design, generating five-second audio-video clips in real time.