MiniMax H3: 2K Video With Native Stereo Audio in One Pass

MiniMax H3 is a 33-billion-parameter dense omni-modal model released on July 31, 2026, and also sold as Hailuo 3.0. It reads text, images, video and audio as a single context and generates 4- to 15-second video at 24 FPS with native 32 kHz stereo audio. The API model ID is MiniMax-H3.

MiniMax H3 is billed per second of output on the MiniMax Open Platform, with 2K priced above the 768p tier; check the live rate card for current figures, since the 768p tier has been gated at times. The Context-IR variant is priced separately by token. On this site, MiniMax H3 generation is credit-metered — 18 credits per second at 768p and 29 credits per second at 2K — with credit packages listed on the pricing page.

Not in the OSI sense. MiniMax published downloadable H3 weights in early August 2026 under a community license. Commercial use requires visible “MiniMax H3” attribution in your interface, model distillation is prohibited, organizations above roughly $20M in annual revenue need separate written authorization, and four territories are excluded outright.

The MiniMax H3 community license excludes the United States, the European Union, the United Kingdom and the Republic of Korea, and hosted access was withheld in those markets at launch — MiniMax cited evolving AI regulation and copyright concerns. Teams in those territories should treat H3 as a model to understand, and use licensed alternatives such as Sora 2, Veo 3.1 or Seedance 1.5 Pro for production work.

Both generate audio with the picture, which puts them in a category of two. MiniMax H3 reaches 2K through its regeneration stage at the lower per-second list rate and accepts richer reference input. Veo 3.1 tops out near 1080p and bills more per second on both of its tiers, but leads on synchronized dialogue, camera predictability, and worldwide availability. For most teams the license question settles it before quality does.

Sora 2 Pro reaches about 25 seconds against H3’s 15-second ceiling and is stronger on physical plausibility, but it generates no audio in the same pass. On list price, MiniMax H3’s 2K rate sits just above Sora 2’s base tier and well below Sora 2 Pro. Once a separate audio stage is priced in, H3 often wins on cost per finished second.

A minimum working set of MiniMax H3 weights is about 42.5 GB, rising to 63.4 GB for both checkpoints and 123.6 GB at full bf16. ComfyUI reports a 12 GB VRAM floor with aggressive offloading, though a 24–32 GB card wants roughly 75 GB of host RAM for INT8 offload. Four H100 80GB cards complete a generation in about 13 seconds; two RTX 5090s take closer to nine minutes for five seconds.

Independent arena tracking in early August 2026 placed MiniMax H3 first in video editing with audio judged, second in text-to-video, and second or third in image-to-video depending on whether audio counted. MiniMax published no official benchmarks of its own, and the confidence intervals overlap its nearest competitor, so read the ranking as close rather than settled.

The Ref2VA route caps images at nine, video clips at three and audio files at three, with a separate combined limit of twelve files — so the category maxima cannot be summed. Reference video and audio each run 2 to 15 seconds, audio requires an accompanying image or video, prompts cap at 7,000 characters, and the request body at 64 MB.

Yes. After sign-in, the Hero workspace can create MiniMax H3 text-to-video, first/last-frame image-to-video, and multimodal Reference-to-Video tasks through KIE. Reference mode accepts images, MP4 or MOV video, and MP3 or WAV audio; generation supports 4–15 second duration and 768P or 2K output. Generation uses this site’s credit balance, while availability and permitted use still depend on your territory and the applicable MiniMax terms.