What Gives MiniMax the Edge in the AI Agent Era?

I recently came across MiniMax's latest business data and found myself staring at the screen for a while. It wasn't because the numbers were exaggerated, but because this set of data finally showed us that, apart from burning money to compete on parameters and vie for C-end traffic, Chinese large model companies have another, more solid way to thrive.

On May 28, this company, which had been listed on the Hong Kong Stock Exchange for just over four months, disclosed that its global enterprise and developer client base had surpassed 1 million, a fivefold increase in half a year, with a global user base of approximately 300 million. Over the past two months, the company's Annual Recurring Revenue (ARR) has grown by over 100%.

Three months ago, at an earnings conference, they mentioned that their ARR in February had exceeded $150 million. A simple calculation reveals that MiniMax's current ARR doubling cycle has been compressed to just 60 days. This speed is unparalleled in the global large model industry, except for Anthropic, which currently has a doubling cycle of about four months.

Many people may not grasp the significance of a 60-day doubling cycle. To put it this way, if a company can maintain this pace, its revenue will grow 64-fold in a year. For a startup founded just over four years ago, this is almost unimaginable growth.

More importantly, this growth is not driven by subsidies or a single hit C-end application. It stems from MiniMax finally identifying its growth engine and maximizing its power.

01 The Slow-Burning MiniMax Suddenly Takes Off

MiniMax has always given the impression of quietly achieving great things.

When Yan Junjie left SenseTime to start his own venture at the end of 2021, ChatGPT had not yet been released, and the entire industry was groping in the dark. At that time, discussions revolved around whether large models could be built, how many parameters were sufficient, and no one dared to say that large models could be profitable, let alone how to make money. Many considered it a capital game that would dissipate once the money was burned.

Yan Junjie and his team chose the quietest path. For more than two years after its establishment, they rarely granted media interviews or made public statements. Most of what the outside world knew about MiniMax came from the few C-end products they quietly launched—Xingye/Talkie and Conch AI. These products, though not heavily promoted domestically, quietly gained popularity overseas, amassing hundreds of millions of users.

This low-key approach led many to underestimate MiniMax's capabilities. It wasn't until January 9 this year, when MiniMax was listed on the Hong Kong Stock Exchange, with its stock price soaring 78% on the first day and a market capitalization nearing HK$90 billion, that people suddenly realized this unassuming company had already emerged as a leader in China's large model industry.

However, skepticism persisted in the market. Many claimed MiniMax relied solely on C-end applications, lacked core technology, and had weak commercialization capabilities. Its total revenue of $79 million in 2025 indeed paled in comparison to OpenAI and Anthropic.

The turning point came in February this year. MiniMax released the M2.5 model, which instantly reached industry-leading levels in programming, tool invocation, and office applications—scenarios with real commercial value. This was followed by the M2.7 model in March, specifically optimized for AI agents, significantly enhancing complex task processing capabilities.

The breakthrough in model capabilities immediately translated into a business boom, with a more critical shift occurring in the revenue structure.

This signifies that MiniMax has undergone a crucial transformation. It is no longer a company reliant on C-end applications but has evolved into a platform-based company centered on B2B and developer services. This transformation not only drives revenue growth but also represents a qualitative leap in revenue quality. API revenue boasts higher gross margins, stronger customer stickiness, and more stable cash flow.

Interestingly, MiniMax's path resembles that of Anthropic. Anthropic avoids advertising, video generation, and hardware, focusing solely on enterprise services and Claude subscriptions. With this "narrow but deep" approach, Anthropic took 15 months to grow its ARR from $1 billion to $45 billion, officially surpassing OpenAI in April this year to become the global leader in ARR for large model companies.

However, MiniMax differs from Anthropic. While Anthropic focuses solely on text models, MiniMax has adhered to a full-modality approach since its inception, simultaneously developing text, voice, image, video, and music modalities. This full-modality capability enables it to offer more comprehensive solutions to clients and leaves ample room for future growth.

02 Developers Are the True Moat for Large Models

As the large model industry has evolved, a consensus has emerged: whoever wins the developers wins the future.

Over the past two years, nearly all large model companies have claimed to be building ecosystems and serving developers. Yet, few have succeeded. Many companies' open platforms are mere facades, with either inadequate model capabilities, incomplete toolchains, or exorbitant prices. Developers leave after a few tries, unable to stay.

MiniMax's open platform was officially launched in the second half of 2025, later than many competitors. Yet, this latecomer achieved 1 million enterprise and developer clients within six months, a growth rate unmatched in China's large model industry.

Many ask how MiniMax managed this. The answer is simple: they genuinely value developers and address their most pressing concerns.

What do developers care about most? Three things: model usability, affordability, and toolchain completeness.

In terms of model capabilities, MiniMax did not pursue "broad but shallow" general capabilities but instead focused on core scenarios most needed by developers: code generation, long-text understanding, tool invocation, and multimodal generation. The performance of M2.5 and M2.7 in these areas has been widely recognized by developers. Particularly in AI agent development, MiniMax is now one of the most popular model providers in China.

Regarding pricing, MiniMax did not blindly engage in domestic price wars but instead pursued a "maximum cost-effectiveness" strategy. By optimizing model architecture and inference engines, they reduced the inference cost per token to the industry's lowest level. Now, the API price for the M2 series models is about one-third of international counterparts, with comparable performance. For developers, this translates to tangible cost advantages.

In terms of toolchains, MiniMax has been more thorough than anyone else. They provide developers with end-to-end tools for model training, fine-tuning, deployment, and monitoring, significantly lowering the barrier to AI application development. In March, they pioneered the adaptation of the open-source framework OpenClaw and launched the MaxClaw cloud-based agent service, enabling individuals and enterprises to quickly build their own digital employees. In April, they introduced MaxHermes, the world's first cloud-based self-evolving AI agent. In May, they upgraded their Agent product to Mavis, supporting multi-agent teamwork.

These products, launched one after another, have formed a complete AI agent development ecosystem. Currently, over 500,000 agents run on MiniMax's platform, covering nearly all industries, including finance, education, healthcare, and e-commerce.

MiniMax also has an unparalleled advantage: globalization. Their C-end products have always targeted global markets, with Xingye/Talkie amassing a large user base and brand recognition overseas. This global advantage extends to the developer ecosystem, with a rapidly growing number of developers from the United States, Europe, and Southeast Asia.

Many claim that the moat for large models lies in computational power, data, or parameters. However, the true moat is developers. As more developers build applications on your platform and their businesses become deeply integrated with your models, switching to another platform becomes difficult. Once this network effect takes hold, it is hard to break.

Anthropic is a prime example. Many enterprises have now deeply integrated Claude into their workflows and would not switch easily, even if OpenAI's models were cheaper. MiniMax is now replicating this advantage, but more thoroughly.

03 The Second Half of the Large Model Race Is Not About Parameters

In 2026, the large model industry officially entered its second half.

The first half was characterized by "speed" and "scale." Companies competed on model size, release speed, and C-end user acquisition. Many rushed to launch half-baked models to gain an edge.

However, the second half is different. Users and enterprises no longer pay for "large parameters" but care about whether models can solve practical problems, improve efficiency, and create value. The focus of industry competition has shifted from model capabilities to scenario implementation and ecosystem building.

This shift is good news for MiniMax, as they have avoided the parameter race and focused on solving practical problems from the start. Their current layout (strategic layout ) aligns perfectly with the rhythm of the second half.

The advent of the AI agent era presents MiniMax's greatest opportunity. At the March earnings conference, Yan Junjie stated that AI would evolve from a tool to a colleague-level collaborator, with explosive growth in software development and workplace productivity, leading to a one- to two-order-of-magnitude increase in platform token demand. This prediction is being validated by the market. More and more enterprises are using AI agents to replace manual labor for repetitive tasks. MiniMax's early layout in the agent field gives them a first-mover advantage.

The explosion of the multimodal market presents another opportunity. Video generation is considered the next killer application after text.

MiniMax's Conch AI has already accumulated rich experience in video generation. Next month, they will release the Conch 3 model, a native "understanding-generation" integrated video large model with a technical architecture comparable to Doubao's Seedance 2.0. If Conch 3 meets expectations, it will bring tremendous growth potential to MiniMax.

Globalization remains MiniMax's long-term advantage. Compared to other domestic large model companies, MiniMax has the highest degree of globalization, with overseas revenue accounting for over 70%. The global AI market is growing rapidly, especially in emerging markets like Southeast Asia and the Middle East, where demand for China's large model technologies is strong. Leveraging its technical advantages and brand recognition, MiniMax is poised to achieve greater breakthroughs in these markets.

Of course, MiniMax faces challenges. Internet giants like ByteDance and Baidu are increasing their investments in large models, boasting more capital, data, and engineering capabilities. OpenAI and Anthropic are also accelerating their entry into the Chinese market, intensifying competition.

Computational power supply is another issue. With rapid business growth, MiniMax's demand for computational power is surging. They now adopt an "asset-light, operations-heavy" model, with hardware held by third parties, and expect most computational infrastructure to be autonomously operated by the end of the year. Simultaneously, they are actively adapting to domestic chips, with full domestic inference capabilities expected to be online by the end of the third quarter this year.

Nevertheless, MiniMax has found its own path. With less than one-tenth of OpenAI's workforce and one-hundredth of its funding, they have achieved growth rates approaching international first-tier levels. This is no small feat.

The large model industry is full of uncertainties, with new stories emerging and bubbles bursting daily. Many companies lose their way chasing trends, forgetting their original purpose. However, MiniMax has maintained its pace, moving forward step by step.

Yan Junjie once said that MiniMax's name comes from the minimax algorithm in game theory, meaning seeking the optimal solution under limited conditions. In the uncertain large model industry, there are no absolutely correct answers, only paths suitable for oneself.

MiniMax has found its engine and its rhythm. This may be the most valuable lesson for Chinese large model companies.

This article is an original piece by Xinmou. For authorization to reprint or business cooperation, please contact us.

— END — For more content, click to follow