In early 2026, an agent tool called OpenClaw went viral globally. The underlying model powering it, M2.5, consumed 1.44 trillion tokens on OpenRouter in a week, topping the platform. That model came from MiniMax, a company listed on Hong Kong's stock exchange just two months earlier. Its stock surged from HK$165 to HK$1,300, pushing its market cap above HK$300 billion.
MiniMax operates with only 385 employees, average age 29. From its founding through September 2025, it spent roughly $500 million. Over the same period, OpenAI burned between $40 billion and $55 billion. With less than 1% of its rival's capital, MiniMax built a full-modality AI platform valued at over $30 billion. The man behind it: Yan Junjie, a PhD from a county in Henan province.
Leaving SenseTime at its peak
Yan started as an intern at Baidu's research institute in 2014. After earning his PhD from the Chinese Academy of Sciences, he joined SenseTime and rose to VP in seven years. In late 2021, he quit just as SenseTime was preparing its Hong Kong IPO. He founded MiniMax in December 2021, nearly a year before ChatGPT launched. This head start helped him raise a $31 million angel round at a $170 million valuation, with investors including miHoYo, IDG, Hillhouse, and Yunqi.
The startup nearly collapsed in March 2023 when Silicon Valley Bank failed, holding all of MiniMax's funds. It survived and two months later closed a $257 million Series A at a $1.157 billion valuation. Alibaba, Tencent, and Sequoia later joined. By its IPO, MiniMax had raised nearly $1.5 billion across seven rounds.
Betting on MoE and extreme efficiency
In summer 2023, Yan gambled by allocating 80% of computing and R&D resources to Mixture of Experts (MoE), a technique then considered cutting-edge but immature in China. His reasoning: serving tens of millions of users with dense models would be prohibitively expensive. In early 2024, MiniMax released China's first MoE large model. The M2.5 model costs just $1 per hour of continuous runtime, one-twentieth the cost of GPT-5. Internally, 80% of code is generated by AI, which staff call "interns" that can directly access code repositories and modify production environments.
MiniMax pursued a full-modality strategy from day one: text, video, voice, and music. Its consumer products include Xingye (China) and Talkie (overseas) for AI companionship, and Hailuo AI for video generation, which led global monthly active users for video generation apps for six consecutive months in 2024. Today it has 236 million users across 200 countries, with overseas revenue accounting for 73%. On the enterprise side, it serves 214,000 business customers and developers, and its models are deployed on Google Vertex AI, Microsoft Azure, and AWS.
Revenue mix puzzle
When MiniMax released the M1 model in June 2025, Yan posted: "For the first time, I feel the mountain can be climbed." But the numbers tell a mixed story. Consumer gross margin is only 4.7% and fell to about 2.1% in Q4, while enterprise gross margin stands at 69.4%. Consumer contributes 67% of revenue but almost no gross profit. Overall gross margin improved from 12.2% to 25.4%, driven by rising enterprise share. This remains an unresolved challenge.
Listed on January 9, 2026, MiniMax saw its stock jump on the first annual report in March. Yan outlined his vision: transform from a large model company to an AI-era platform company, defining platform value as "intelligence density × token throughput." His competitors are Google and OpenAI, with resources tens of times larger.
The U.S. chip embargo forced Chinese AI firms onto a different path. Limited to less powerful chips, MiniMax innovated with Mixture of Attention architectures, linear attention, and the CISPO algorithm. With 385 people and $500 million, it's up against OpenAI's 4,000+ employees and plans to spend $600 billion on compute by 2030. The race is far from over.