MiniMax unveils M2.5 model to rival Anthropic, OpenAI

SHANGHAI: Chinese artificial intelligence unicorn MiniMax has released its latest large language model, M2.5, intensifying the fierce competition within China’s tech sector just ahead of the Spring Festival holiday.

The company asserts that the new advanced model is developed particularly for “real-world productivity.” It offers a captivating mix of high performance and low operational costs that rivals American AI models like OpenAI and Anthropic.

The M2.5 release comes as global investor interest in Chinese AI startups increases, looking beyond traditional tech giants to capitalize on quick innovation.

MiniMax’s M2.5 model has maintained a consistent, streamlined size of 230 billion parameters across its iterations. Despite this, the company reported that M2.5 performs comparably to leading global models in critical areas such as coding and search.

MiniMax emphasized the model’s cost-efficiency, marketing it as the “first frontier model where users do not need to worry about cost.” The company claims the model delivers “intelligence too cheap to meter,” capable of running continuously for an hour at 100 tokens per second for just US$1.

To demonstrate its practical application, MiniMax has integrated M2.5 into its “MiniMax Agent,” an AI intern tool launched last month. The company reports that the agent has already autonomously completed 30% of internal tasks across its research, sales, and HR departments.

On an internal model inspired by OpenAI’s “GDPval,” which tracks performance across real-world occupations, M2.5 gained a score of 59%, a significant jump from its predecessor’s 24.6%. Though it still trails Anthropic’s Claude Opus 4.6, which scored 73.5%.

The launch follows a busy week for the industry, with domestic rival Zhipu AI recently rolling out its massive 744-billion-parameter GLM-5 model. The increasingly AI race mirrors the strategy set by DeepSeek last year, as companies race to unleash major updates before the holidays.

About MiniMax

MiniMax is riding a wave of momentum financially. It was founded in late 2021, and HK$4.8 billion (US$614.1 million) was raised in its Hong Kong IPO last month, with stock prices rising over 70% since the debut.

In a recent opinion piece, CEO Yan Junjie observed that “limited computing power” has not hindered Chinese companies from becoming global players. The company’s prospectus supports this assertion, revealing that applications like Talkie and Hailuo generate over 70% of its revenue from international markets.

MiniMax intends to open-source the model by releasing its weights, although it has not specified the semiconductor chips used for training.