en.Wedoany.com Reported - Chinese AI startup MiniMax Group is developing a large language model with 2.7 trillion parameters, which could be released as early as the third quarter. This would be the largest open-weight AI model released by a domestic company and could also become one of the largest models globally.
The global push for trillion-parameter AI models stems from the growing demand for autonomous systems capable of complex reasoning. Analysts believe that this scale threshold is crucial for building systems that can execute multi-step operations without human intervention. A person familiar with the matter told Reuters, speaking on condition of anonymity as the information has not been made public. MiniMax declined to comment.
The plan to develop this massive model (details first reported by The Information) comes as cheap models based on open-source technology from Chinese AI providers gain traction in the U.S. and other markets, serving as low-cost alternatives to proprietary American systems. MiniMax's competitors include companies such as Z.ai and DeepSeek. Open-weight models allow users to download, run, and customize the underlying system, unlike proprietary closed-source models.
The source also added that MiniMax will launch its frontier-level multimodal video generation model H3 later this month.
Founded in 2021, MiniMax is a rising star in China's AI industry. In January, the company raised HK$4.8 billion (US$614 million) in its initial public offering in Hong Kong and plans a secondary listing on Shanghai's STAR Market. Currently, Meituan's LongCat-2.0 and DeepSeek's V4-Pro lead China's AI industry with a total of 1.6 trillion parameters, while several other domestic competitors have also crossed the trillion-parameter threshold.
Although standard generative chatbots excel at processing short texts, they have mathematical limitations when it comes to making independent long-term decisions.
This large-scale architectural expansion relies heavily on mixture-of-experts engineering techniques to balance intelligence levels with operational costs. By organizing the model into specialized subnetworks, developers can build vast trillion-parameter databases, activating only a fraction of their capabilities for each query. This approach enables users to obtain deeply specialized domain knowledge—including complex legal regulations and rare software defects—at the high speed and low cost of a medium-sized system.
