MiniMax unveiled its open source H3 multimodal generation model, offering 2K video generation with native audio at one-twelfth the API cost of ByteDance’s Seedance 2.5.
MiniMax officially released its next-generation universal multimodal AI generation model, designated as H3, on August 3, 2026. MiniMax announced that it will open-source the model weights for H3 within a few days, enabling enterprise developers to execute local, private deployments and mitigate critical data sovereignty risks.
The H3 model is directly benchmarked against ByteDance’s closed-source Seedance series (specifically Seedance 2.5) in the rapidly growing AI video generation market. MiniMax priced H3’s API at CNY¥0.8 per second (approximately USD$1.95 for a 15-second 2K video), which is one-twelfth the cost of Seedance 2.5 (CNY¥10 per second).
H3 generates up to 15-second video sequences at default 2K resolution with native dual-channel audio, cross-modal composite referencing, and advanced text rendering capabilities. The company’s proprietary VAE architecture (H3-VAE) compresses video sequence lengths fourfold, substantially lowering training and inference compute costs to enable its disruptive low-price strategy.
According to independent Artificial Analysis evaluations, H3 ranks first globally in video editing capability, second in text-to-video, and third in image-to-video capabilities. Following the pre-market announcement on August 3, 2026, MiniMax Group Inc. shares (0100.HK) surged by over 10% on the Hong Kong Stock Exchange to close at HKD$249.40.
