MiniMax Plans to Leapfrog AI Rivals with Giant 2.7 Trillion Parameter Model
By Reuters | 08 Jul, 2026

The Chinese AI startup is planning to launch what may be the largest open-weight AI model to tap the growing demand for autonomous systems capable of complex reasoning.

(Image from MiniMax homepage promotion)

Chinese AI startup MiniMax Group is working on a large language model with 2.7 trillion parameters, a person with direct knowledge told Reuters, the largest open-weight AI model released by a Chinese firm and possibly the largest in the world.

The global push toward trillion-parameter AI models is being driven by growing demand for autonomous systems capable of complex reasoning. Analysts say the threshold is critical to building systems that can execute multi-step operations without human intervention.

The new model could be released as early as the third quarter, the person said, declining to be identified as the information is not public. The company declined to comment.

LOWER-COST ALTERNATIVES TO U.S. SYSTEMS

Plans for the colossal model — details of which were first reported by The Information — come as cheaper, open-source-based models from Chinese providers such as MiniMax, Z.ai and DeepSeek are gaining traction in the U.S. and other hubs as lower-cost alternatives to proprietary U.S. systems.

Open-weight models allow users to download, run and customise the underlying systems, unlike proprietary, closed-source models.

MiniMax will later this month launch H3, its frontier-level multimodal video generation model, the source added.

Founded in 2021, MiniMax is an upstart in China's AI industry, raising HK$4.8 billion ($614 million) in its Hong Kong initial public offering in January. It is planning a second listing in Shanghai's tech-heavy exchange, the STAR Market.

Meituan's LongCat-2.0 and DeepSeek's V4-Pro lead China's AI industry with 1.6 trillion total parameters, while several other domestic rivals have passed the trillion-parameter threshold.

While standard generative chatbots excel at short-form processing, they hit mathematical limitations when tasked with independent, long-horizon decision making.

MIXTURE OF EXPERTS ENGINEERING

This massive architectural expansion relies heavily on Mixture of Experts engineering to balance intelligence with operational costs.

By organizing a model into specialized subnetworks, developers can build vast, trillion-parameter databases that only activate a small fraction of their capacity per query.

The approach gives users access to deeply specialized domain knowledge, including complex legal codes and rare software bugs, at the high speed and lower cost of mid-sized systems.

(Reporting by Reuters staff; Editing by Eduardo Baptista, Anil D'Silva and Jan Harvey)