MiniMax H3

Open Weights

MiniMax

MiniMax's open-weight omni-modal generation system. It takes a unified multimodal context of text, images, video and audio and produces video with native 32 kHz stereo audio at 24 fps, 4 to 15 seconds long, with the shorter side set to 768 pixels by default. The open release is H3-Base: two CFG-distilled checkpoints, FL2VA for text and first/last-frame input and Ref2VA for reference images, video and audio, built on a 33B-parameter dense single-stream Omni Transformer that draws hidden states from a Qwen3-VL-32B encoder. 2K output requires the hosted H3-Regenerate-2K module, which is not open sourced.

Strengths

Jointly generates video and native 32 kHz stereo audio rather than dubbing a separate audio pass
Omni-reference input accepts up to 9 images, 3 video clips and 3 audio clips, at most 12 files per request
Output durations of 4 to 15 seconds across aspect ratios including 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16 at 24 fps
Stable dialogue support for 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish
Full 33B Omni Transformer weights released, with AdaLN branch parameters that need not be loaded for inference-only deployment
First-party serving recipes for SGLang, vLLM, Diffusers and ComfyUI, plus reproducible 768p request scripts

Caveats

The licence's "Excluded Territories" are the European Union, the United Kingdom, the Republic of Korea and the United States of America, so it grants no rights in those regions without a separate application
Only H3-Base at 768p is open: H3-Regenerate-2K "is not yet open-sourced" and H3-Context-IR "is not included in this open-source release", both reachable only through MiniMax's hosted API - so the headline 2K output cannot be reproduced locally
Sparse attention, which the model was trained with, is missing: "The initial open-source release provides inference with full attention only"
No numeric benchmark scores are published as text on the model card or the launch post, and the price claims are purely relative ("less than a third of mainstream models") with no rate given
Inputs and enhanced prompts pass through automated moderation that "may be blocked" and which MiniMax says "cannot eliminate false positives or false negatives"

Vendor: MiniMax
License: MiniMax H3 Community License Agreement
Modalities: video
audio

Capabilities

Vision

Audio

Video

Tool Use

Resources

API Documentation Hugging Face GitHub

Pricing

Weights download free under the MiniMax H3 Community License Agreement, whose "Applicable Territory" is "worldwide, excluding the Excluded Territories" - defined as "the European Union, the United Kingdom, the Republic of Korea and the United States of America" - with an application form offered for those regions. Hosted access is metered through the MiniMax platform video-generation-v2 API (model ids MiniMax-H3 and MiniMax-H3-Max) and the Hailuo AI app; MiniMax states its price only in relative terms and publishes no per-second rate as text.