MiniMax AI Co., Ltd. is a Chinese artificial intelligence model developer registered in the People's Republic of China and headquartered in Shanghai. It was founded in December 2021 by a group of former SenseTime employees.
MiniMax AI quickly established itself as a technology leader, developing efficient proprietary models and popular applications based on them, such as the chat application Talkie and the video generation platform Hailuo AI. Within the research community, the company gained recognition for its architectural solutions for processing ultra-long contexts, first introduced in the MiniMax-01 series of open models. This refers to the Lightning Attention mechanism integrated with the Mixture-of-Experts (MoE) architecture. This architecture enabled the training of models with contexts of up to 1 million tokens and extrapolation to 4 million tokens during inference. The next step was the release of the open model MiniMax-M1, which introduced the concept of "budgeted thinking" with explicitly set limits (40K/80K tokens). A new RL algorithm called CISPO was developed for its training, significantly accelerating the learning process.
Notable product developments by the company include Video-01/Hailuo-02—text-to-video generation services with resolutions up to 1080p, Speech 2.5—a rapid voice cloning technology supporting multiple languages, and the Realtime API—a low-latency interface for native multimodal speech and text processing. However, since 2025, MiniMax AI has also actively begun developing its open-source direction, releasing models with open weights. Among the first open releases were: MiniMax-Text-01, MiniMax-VL-01 (under the MIT license), MiniMax-M1-40K, and M1-80K (under the Apache 2.0 license)—models that generated genuine interest and positive feedback from the community.