MiniMax H3 AI Video Generator

MiniMax H3 is an omni-modal video model released on July 31, 2026 that generates 5 to 15 second clips at up to 2K resolution with native stereo audio. It reads text, images, video, and audio in a single context rather than handling each input type separately. A request can combine a written prompt of up to 7,000 characters with nine reference images, three video clips, and three audio tracks. The model also performs instruction-based editing, changing one element of a finished clip while leaving everything else stable.

You can start free. New accounts receive 20 credits with no card required, which is exactly one 5-second clip at full 2K with native audio — the same model paying customers use, no watermark and no quality cap. Spend it on one 2K render or two 768p drafts, whichever tells you more. When you need more, you buy a credit pack starting at $9.90. There is no subscription to cancel.

Clips run 5 to 15 seconds at up to 2560×1440 — native 2K — at 24 frames per second. On a 21:9 frame that comes out around 2976×1248. Six aspect ratios are available: 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16. Longer durations and higher resolutions cost more credits, so a 5-second 768p test is cheaper than a 15-second 2K final.

Yes. Every generation returns stereo audio built in the same pass as the picture — dialogue, ambient sound, music, and effects, synced to the action. This is unusual: most video models return silent clips. You can also attach a voice sample as a reference and have a character speak new lines in that voice. If you do not describe the sound in your prompt, H3 infers it from the scene, so it is worth stating explicitly.

Twelve per request: up to nine images, three video clips, and three audio tracks. You tag each one in the prompt so the model knows its role — a face for identity, a clip for motion, a still for color, a sample for voice. Supported formats include JPG, PNG, WEBP, and HEIC for images, H.264 and H.265 for video, and WAV and MP3 for audio.

Depends where you are. MiniMax published H3’s base weights on August 3, 2026, and the model runs through ComfyUI, vLLM, SGLang, and Hugging Face diffusers. But the MiniMax H3 Community License excludes local deployment in the United States, the European Union, the United Kingdom, and South Korea. There is also a hardware floor: this is a 33-billion-parameter model decoding video and audio together, and consumer GPUs generally do not have the VRAM for it at full precision. Running it in the browser avoids both problems.

Output is charged per second: 4 credits at 2K, 2 credits at 768p. A 5-second 2K clip is 20 credits, which lands between $1.20 and $2.00 depending on which credit pack you bought. Text prompts, audio, and up to five reference images cost nothing extra. Reference videos are charged on their own duration, so a long clip attached to a short render can cost more than the render. Failed or rejected generations are refunded automatically. Full breakdown on the pricing page.

The main gap is resolution: H3 renders at 2K, Seedance 2.0 at 720p or 1080p. Otherwise they are close — both cap around 15 seconds, both generate audio with the video, and both accept a similar set of reference files. Seedance 2.5, released in June 2026, goes further than H3 on length and resolution with 30-second clips and 4K output. H3’s remaining advantage over both is that its weights are public.

Videos generated on a paid plan come with commercial use rights, which covers advertising, e-commerce listings, social marketing, brand films, and client work. H3 renders brand text, packaging detail, and UI elements accurately enough for that to be practical rather than theoretical. Check the license page for the full terms before using output in a paid campaign.

Name the subject, the action, the camera, the lighting, and the sound. H3 responds to real camera language — dolly in, handheld, hard cut, rack focus, title card — and those terms are more precise than describing the effect you are after. Prompts can run to 7,000 characters, so a structured shot list outperforms a one-line description. Every reference you attach is one less thing you have to write.

Turbo is a faster generation mode that trades some render quality for speed and lower credit cost. It is the right choice while you are still iterating on framing and timing — run Turbo until the shot is right, then re-render the final at full quality. Credit cost differs between modes, so check the pricing page before batching a large job.