MiniMax released H3, its first open-source multimodal generation model, on July 31, combining 2K video output, native audio-video generation, and pricing at roughly one-third of comparable flagship models1. The launch arrives alongside ByteDance's rollout of its Seedance 2.5 model, intensifying competition in Chinese AI video generation3.
H3 ranks first globally in video editing on the Artificial Analysis benchmark leaderboard. However, the model trails Google's Gemini Omni Flash in text-to-video tasks and also ranks behind ByteDance's Seedance 2.0 in that category. The model supports 15-second audio-visual generation at 2K resolution, outputting native dual-channel audio-video from text, image, video, and audio inputs4.
A core technical differentiator is V2V motion transfer — the ability to migrate motion features from a source video to a target video for controllable multimodal content editing. MiniMax positions H3 for commercial use cases including advertising, brand content, e-commerce, product design, UI/UX, and gaming.
Pricing is set at 0.8 yuan per second of generated video. The cost advantage stems from systematic optimization: a high-compression tokenizer reduces the number of video generation tokens required, while heterogeneous training tuning, load balancing, and GPU utilization efficiency lower both training and inference costs.
The open-source strategy is central to the release. MiniMax plans to open H3's model weights within days, subject to regulatory compliance5. The company said the move is intended to break closed-source "dominance" through open weights and competitive pricing. MiniMax also said chip manufacturers and developers can participate in adaptation and optimization to reduce usage costs and expand application scope.
MiniMax shares rose 14% on the announcement, reversing a prior decline, as investors backed the cost-control strategy.
ANALYSIS The H3 release sharpens the competitive dynamic in Chinese AI video generation, with MiniMax leveraging open weights and aggressive pricing against ByteDance's closed-source Seedance line. The benchmark results present a mixed picture: H3 leads in video editing but lags in text-to-video, suggesting MiniMax is carving out a niche in controllable editing rather than competing head-to-head across all generation tasks. The pricing and open-weight combination lowers adoption barriers for enterprise developers, particularly in cost-sensitive commercial production workflows.