MiniMax H3 is a new omni-modal AI video model generating 15-second 2K clips with native stereo audio. Here's what HK creators need to know.
MiniMax just shipped a new AI video model that folds text, images, video, and audio into one generation pipeline. MiniMax H3 generates 15-second 2K clips with native stereo audio — no separate audio track needed.
Launched on July 31, 2026, H3 is available through the MiniMax API and the Hailuo AI app. For Hong Kong creators, this is a notable step forward in AI video production.
What Makes MiniMax H3 Different
Most AI video tools split tasks across separate models — one for text-to-video, another for image-to-video, another for motion reference. MiniMax H3 does it all in one pass. The model accepts text, images, video clips, and audio files as a unified input and returns video with stereo sound baked in.
MiniMax's example prompt illustrates this: "Reference the camera movement from Video 1, have the character in Image 2 sing, match the vocals to Audio 3." One prompt, one generation, one output with everything synced.
You can supply up to 9 reference images, 3 video clips, and 3 audio clips per request. Output is 2K resolution at 4-15 seconds.
2K at Lower Cost
The key technical innovation is the H3-VAE tokenizer, which delivers 4x the effective sequence length of previous architectures. This makes native 2K generation economically viable — MiniMax claims the per-second price at 2K is less than a third of mainstream models.
Reports put the pay-as-you-go rate at roughly $0.13 per second for 2K, or about $1.95 for a full 15-second clip. For Hong Kong agencies producing high volumes of short-form social media content, that pricing is competitive.
The H3-Omni Transformer architecture separates understanding and generation workloads, achieving nearly 30% better training throughput.
Detail Preservation for Brand Content
Instead of using a separate upscaling module, H3 regenerates its own low-resolution output in-context — re-reading the original input to recover small text and fine detail. This matters for brand content where logos, labels, and packaging text must stay crisp.
According to Artificial Analysis, H3 leads the video editing category, though it trails Google's Gemini Omni Flash in text-to-video and ByteDance's Seedance 2.0 in image-to-video.
What This Means for Hong Kong Creators
For ad agencies producing product videos, social media content, and brand storytelling, the ability to generate clips with matching audio in one go speeds up production significantly. MiniMax positions H3 for advertising, e-commerce, branding, gaming, and film pre-visualisation.
Open weights are promised "in the coming days," but for now, access is through the API only. Cooly Studio users can integrate the MiniMax H3 API alongside other tools in their AI production pipelines.
Frequently Asked Questions
Q: How does MiniMax H3 compare to Seedance 2.5? A: Seedance 2.5 generates up to 30-second clips with audio, while H3 tops out at 15 seconds but delivers native 2K. H3 leads in video editing benchmarks.
Q: Is MiniMax H3 available in Hong Kong? A: Yes. The model is live on the MiniMax platform API (model ID: MiniMax-H3) and in the Hailuo AI app, both accessible from Hong Kong.
Q: How much does MiniMax H3 cost? A: Roughly $0.13 per second for 2K, or about $1.95 for a 15-second clip — under a third of mainstream 2K alternatives.
Q: Can I use MiniMax H3 with Cooly Studio? A: Cooly Studio supports a range of AI video models. The MiniMax H3 API can be integrated into custom workflows alongside other tools in the cooly.ai ecosystem.
Q: What input formats does H3 accept? A: H.264/H.265 video, JPG/PNG/WEBP/HEIC/HEIF images, and WAV/MP3 audio. Max file sizes: 50 MB per video, 30 MB per image, 15 MB per audio.