Another Major Chinese AI Model Goes Open Source: MiniMax H3 Tops Global Audio-Video Editing Benchmark, 16 Chip and Platform Partners on Day One
Text, image, audio, and video — all in one model.
On August 3, MiniMax officially open-sourced its next-generation universal video model, MiniMax H3.
MiniMax H3 is a universal omni-modal generation system capable of understanding multimodal context composed of text, images, video, and audio. It generates videos up to 15 seconds long at up to 2K resolution with native stereo audio.
MiniMax H3 was first released on July 31. On the Artificial Analysis audio-video editing leaderboard, MiniMax H3 currently ranks first with an Elo score of 1,130 — ahead of Gemini Omni Flash, HappyHorse-1.0, Wan 2.7, and other leading video models worldwide.
The H3 system consists of three modules: H3-Context-IR, H3-Base, and H3-Regenerate-2K.
Developers can download MiniMax H3 directly from Hugging Face. H3-Base currently supports deployment via inference frameworks and workflows including SGLang, vLLM, diffusers, and ComfyUI.
Alongside the open-source release, 16 ecosystem partners have completed adaptation support on day one.
These include chip manufacturers such as Huawei Ascend, Moore Threads, Metax (Muxi), Hygon (Haiguang), Kunlun Chip, Iluvatar CoreX (Tianshu Zhixin), Biren Technology, AMD, and Intel; developer communities and cloud inference platforms including Hugging Face, ModelScope, ComfyUI, RunningHub, and fal; and inference frameworks such as vLLM-Omni and SGLang.
Open-source repository:
huggingface.co/MiniMaxAI/MiniMax-H3
Model demo and API:
- H3-2K direct output: platform.minimaxi.com/docs/api-reference/video-generation-v2-create
- H3-Context-IR: platform.minimaxi.com/docs/api-reference/video-generation-v2-h3-context-ir
- H3-Regenerate-2K: platform.minimaxi.com/docs/api-reference/video-generation-v2-regeneration
- MiniMax Hub: hub.minimaxi.com
- Hailuo AI: hailuoai.com
- Wavespeed AI MiniMax H3
We tested MiniMax H3 hands-on. The model can already handle a variety of video generation tasks and performs well in terms of visual naturalness, shot composition, and overall style consistency.
There is room for improvement in photorealism and shot sequencing.
For example, we asked the model to generate a 15-second aerial landscape video. The resulting footage featured realistic snow-capped mountains, grasslands, and other natural scenery, with natural-looking color, lighting, and aerial camera movements. The long take maintained good visual coherence overall. However, temple structures in the footage still showed noticeable AI-generated artifacts.
We then asked MiniMax H3 to generate a promotional ad video featuring a milk tea brand and a gaming IP crossover. The generated video contained six shots with a clear narrative sequence and a generally unified visual style and advertising tone. However, in terms of multi-shot sequencing, the model produced one noticeable repetition: the action of a staff member handing a cup of milk tea to a customer appeared twice in succession.
1. Up to 15-Second 2K Video with Synchronized Stereo Audio
MiniMax H3 is a universal omni-modal generation system. In terms of output specifications, H3 can generate videos from 4 to 15 seconds, supports aspect ratios including 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16, outputs at 24 FPS, and can simultaneously generate 32kHz stereo audio.
The model generates video with a default short-side resolution of 768 pixels, which can be upscaled to 2K via H3-Regenerate-2K. H3 stably supports 11 languages including Chinese, English, Japanese, Korean, French, and German, with varying degrees of support for additional languages.
H3 includes two versions — FL2VA and Ref2VA — designed for first-and-last-frame generation and omni-modal reference generation, respectively.
H3-Base-FL2VA operates in first-and-last-frame mode and accepts up to two input images. Without image input, the model performs text-to-video generation. With a single first-frame or last-frame image, it generates a corresponding video. With two images, it generates the intermediate video content between the specified start and end frames.
H3-Base-Ref2VA supports omni-modal reference mode with up to 9 input images; up to 3 video clips (each 2–15 seconds, total not exceeding 15 seconds); and up to 3 audio clips (each 2–15 seconds, total not exceeding 15 seconds, must be paired with an image or video input). The combined total of image, video, and audio files cannot exceed 12.
2. Three-Module System for Audio-Video Generation with Context-Aware 2K Upscaling
The H3 system comprises three modules: H3-Context-IR (understanding and organizing multimodal instructions), H3-Base (generating audio and video), and H3-Regenerate-2K (producing 2K output).
H3-Context-IR is a managed preprocessing and orchestration system that understands the relationships between text, images, audio, and reference videos, as well as how these materials relate to the intended output. Its internal workflow includes instruction parsing, cross-modal association, temporal understanding, and complex logical reasoning.
H3-Context-IR relies on a multi-stage workflow with multiple managed models and services. This module is not yet open-sourced. MiniMax currently provides corresponding APIs and prompt guidelines, enabling developers to build their own multimodal preprocessing systems.
- Video prompt writing guide: huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md
- Full reference mode output guide: huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md
H3-Base handles the actual audio-video generation. Text is encoded by H3-Encoder; visual inputs are processed by both H3-Encoder and H3-VisualVAE; audio is encoded by H3-AudioVAE. Information from different modalities is then organized into a unified sequence and fed into the H3-Omni-Transformer.
For 2K resolution output, H3 does not use a traditional dedicated super-resolution module. Instead, H3-Regenerate-2K has the base model re-generate from the original text and multimodal references, using the already-generated 768p video as a guide. This approach re-leverages the original context, helping restore fine details such as small text and intricate textures that are difficult to reconstruct from low-resolution footage alone. This module is also not yet open-sourced; developers can replicate the full 2K workflow via the official API.
3. Local Deployment Supports 768p Audio-Video; Full 2K Workflow Requires API
MiniMax provides developers with two verification paths: H3-Base local deployment and the full 2K workflow.
H3-Base is published as two separate model repositories (FL2VA and Ref2VA), each containing processors, tokenizers, a text encoder, the Omni Transformer, Visual VAE, and Audio VAE inference components. These can be deployed via SGLang, vLLM, diffusers, and ComfyUI, with support for multi-GPU parallel inference.
With H3-Base deployed locally, developers can generate audio-video at a short-side resolution of 768 pixels. The FL2VA version supports text-to-audio-video and first-and-last-frame generation, while the Ref2VA version supports tasks such as character and scene preservation, action and lip-sync editing, and voice reference using joint reference images, videos, and audio.
The full 2K workflow combines local models with the official API: H3-Context-IR first interprets and expands user input, then the locally deployed H3-Base generates 768p audio-video, and finally H3-Regenerate-2K re-generates 2K video using the original context. MiniMax also provides reproducible examples for text-to-audio-video, first-frame-to-audio-video, and omni-modal reference generation — complete with request parameters, sample code, and reference outputs available on the Hugging Face model page.
4. Conclusion: Omni-Modal Video Generation Moves Toward an Open Ecosystem
In terms of real-world performance, MiniMax H3 can already handle diverse generation tasks including natural landscapes and commercial advertising, demonstrating a high level of competence in visual naturalness, shot composition, and style consistency.
With multiple chip manufacturers, developer communities, cloud inference platforms, and inference frameworks completing adaptation support simultaneously, MiniMax H3's open-source release not only opens up the model's capabilities but also establishes an end-to-end ecosystem pipeline from model download to local deployment to application development.
Competition among omni-modal video models is expanding beyond visual quality alone into audio generation, complex instruction understanding, and industrial ecosystem building.
Top comments (0)