China's large models are creating a fierce storm in the global developer community, named "cost-effectiveness" and "engineering."

On March 3, 2026, according to the latest data from the world's largest API aggregation platform OpenRouter, domestic large models have shown unprecedented dominance: MiniMax M2.5 leads the world in call volume, followed by Kimi K2.5 and GLM-5. This "domestic top three" ranking marks that Chinese manufacturers have surpassed overseas competitors through their extreme efficiency advantage in AI application deployment and developer preferences.

7 Days Consuming 3.07 Trillion Tokens: MiniMax's "Brutal" Growth

Since its release on February 13, MiniMax M2.5 achieved a call volume of 3.07T (3.07 trillion) Tokens within just one week. This model, known as "the world's first native design for agent scenarios," precisely addresses developers' core pain points:

  • Agent Fundamentalism: It is not only a chat interface but also a backend engine capable of executing complex programming and automated workflows.

  • Price Killer: The extremely low call cost allows developers to build high-frequency-triggered AI applications without any psychological burden.

  • Financial Miracle: Thanks to the explosion of M2.5, MiniMax's ARR (Annual Recurring Revenue) surged to $150 million in February, completing its commercialization goals for the year ahead of schedule.

The Art of Trade-off: Why Only 10B Activated Parameters?

In an industry where others often pursue billions or trillions of parameters, MiniMax M2.5 has only 10B activated parameters. The official explanation states that this is to balance reasoning efficiency and actual output. By optimizing the MoE (Mixture of Experts) architecture, MiniMax has proven that a small size can deliver programming logic comparable to top models. However, the relatively limited knowledge capacity remains a major drawback reported by users.

Next Week's Duel: M3.0 vs DeepSeek V4

The competition in China's large model market has entered a highly intense phase with weekly updates:

  • DeepSeek's Surprise Attack: DeepSeek V4 has been confirmed to be officially released next week, with the global community holding its breath for its latest cost-reduction and efficiency-enhancement solutions.

  • MiniMax 3.0 Preview: XiYu Technology plans to launch the MiniMax M3 Series this year.

  • Morgan Stanley Highly Optimistic: Morgan Stanley's latest report points out that M3.0 will not only introduce more pre-training and architectural innovations to solve the knowledge capacity issue but also support powerful multimodal understanding. Its overall performance is expected to truly rival the world's top flagship models.