MiniMax H3 Review

MiniMax H3 is a video generation model from MiniMax, released on July 31, 2026. MiniMax describes it as a general-purpose multimodal generation model: one prompt can mix text with images, video clips and audio, and the relationships between them are described in plain language. It generates 4 to 15 second clips at 24 fps with 32 kHz stereo sound (voice, effects and music are generated together), in aspect ratios from 21:9 to 9:16. Beyond text-to-video and image-to-video it supports first-and-last-frame generation, an omni-reference mode (up to 9 images, 3 videos and 3 audio clips, 12 files in total), video-to-video motion transfer, and editing that replaces, adds, removes, relights or revoices elements of an existing clip. MiniMax says dialogue works reliably in 11 languages, including English, Chinese, Spanish, French, German, Japanese and Korean.

The full system has three parts: H3-Context-IR, which rewrites the multimodal prompt; H3-Base, a 33B-parameter transformer that renders at 768p; and H3-Regenerate-2K, which re-renders the result at 2K. You can use it in MiniMax's Hailuo AI web app and the MiniMax Hub desktop app, through the MiniMax API (model ID MiniMax-H3, plus Highspeed, Lite and Max variants), or by downloading the H3-Base weights from Hugging Face under the MiniMax H3 Community License. The open release is the base model only: Context-IR and the 2K regenerator are API-only for now, and the published example runs on 4 GPUs.

The MiniMax API is pay as you go, billed per second of output video. Text-to-video, image-to-video, first-and-last-frame and reference tasks cost the same for a given model and resolution:

The first 5 input images per task are free (first 2 on Max), and audio input is free. Other API items:

In practice: a 5-second 768P clip from text costs 5 x $0.08 = $0.40, a 15-second 2K clip costs 15 x $0.13 = $1.95, and a 5-second Lite 480P draft costs 5 x $0.025 = $0.125. Editing a 10-second clip at 768P adds 10 x $0.08 = $0.80 for the input video on top of the output.

In the Hailuo AI app, H3 is paid for with plan credits; MiniMax's H3 page estimates 2K output at about $0.073 to $0.120 per second (about $0.37 to $0.60 for a 5-second clip without video input), depending on the plan. Hailuo plan prices were not readable on the subscription page today.

The open weights are free to download, but the license excludes the United States, the European Union, the United Kingdom and South Korea (users there must ask MiniMax for a license), requires a visible "MiniMax H3" label in commercial products, and needs written authorization above $20 million in yearly revenue.

Prices checked on the vendor's pricing page on November 10, 2026. See MiniMax API pricing and MiniMax H3 on Hailuo AI.

What MiniMax's blog, model card and app page list for H3:

H3 suits developers and studios that want a video model with sound in one pass and per-second API pricing, especially for editing existing footage or combining several image, video and audio references in one prompt. The Lite and Highspeed variants make cheap drafts possible before a 2K render. Teams outside the license's excluded territories can also fine-tune or self-host the base model, if they have multi-GPU hardware.

If you are in the US, EU, UK or South Korea and need self-hosting, the H3 weights are not licensed for you; use the API or pick another model. For clips longer than 15 seconds, Kling generates up to 30 seconds in one pass. Compare Seedance 2.0 and Google Veo on audio-video quality, and Runway if you want a full editing suite around the model.

Coverage so far is from press rather than long-term users. The Batch from DeepLearning.AI (August 2026) reports H3 placing first on Artificial Analysis's video editing leaderboard and near the top for text-to-video and image-to-video, but argues the release is "only minimally open": only the base model is downloadable and the license shuts out the US, UK, EU and South Korea. The South China Morning Post frames H3 as a low-price challenge to ByteDance's video models. MiniMax's own H3 page cites Artificial Analysis Video Arena rankings as of July 31, 2026 (second for text-to-video with audio, first for video editing with audio); those leaderboards change often, so check them before relying on a rank.