AI News (2026/7/31): MiniMax H3 Full-Modal Generation Model Officially Released: Native Stereo Audio and Video, Up to 15 Seconds of 2K Resolution, Open Source Planned

Model Overview

MiniMax has officially launched the MiniMax H3 full-modal generation model. Described as a general-purpose model for full-modal generation, it can uniformly understand and natively generate text, images, videos, and audio, achieving unified multi-modal generation. The model directly outputs videos with native stereo audio, supporting up to 15 seconds of 2K resolution (2560×1440). It also supports commercial scenarios such as instruction following, brand information presentation, and video-to-video motion transfer (V2V Motion Transfer).

Core Technologies and Capabilities

MiniMax H3 incorporates several self-developed technologies to achieve unified generation across all modalities:

• Contextual Omni Representation: The model simultaneously understands the relationship between multi-modal context and the target output. Each inference on a piece of content consumes approximately 100,000 tokens, generating an average of 4,000 tokens of detailed description. Language serves as a bridge, connecting and interpreting across modalities, allowing tasks to be unified in an open descriptive format.
• H3-VAE (Efficient Visual and Audio Encoder): Compared to previous tokenizers, compression efficiency has improved significantly, offering a 4x gain in sequence length. This reduces training and inference costs while supporting native 2K resolution output.
• H3-Omni Transformer (Heterogeneous Training Architecture): With the introduction of multi-modal context, sequence length variance has increased threefold, and the computational load for understanding and generation has become heterogeneous. H3 uses a heterogeneous training architecture to fine-tune hardware utilization, achieving an end-to-end training throughput increase of nearly 30%.

Usage and Cost-Effectiveness

Users can access the H3 model through the MiniMax official website, API, or the MiniMax Hub creation platform. The price per second at 2K resolution is less than one-third that of mainstream models, and at 768P it is less than half, making it highly cost-effective. The model was designed from the outset to support multiple domestic chips and localized deployment.

Regarding open source, the company has announced plans to open-source the model weights in the near future, supporting domestic chip compatibility and customized development. Third-party platform fal.ai has already listed it as an open-weights model, allowing users to experience or deploy it directly.

Sources and Project Links

• MiniMax Official Website: https://www.minimax.io/
• MiniMax Open Platform: https://platform.minimax.io/
• MiniMax Hub (Creation Platform): http://hub.minimax.io/
• fal.ai Model Page: https://fal.ai/minimax-h3
• Official Twitter Announcement: https://x.com/MiniMax_AI/status/2083008095488516262
• IT Home Report: https://www.ithome.com/0/983/957.htm
• China.com Report: https://m.tech.china.com/article/20260731/202607311930284.html
• Tencent News: https://news.qq.com/rain/a/20260731A03M1400