Yan Junjie once had a problem that no benchmark could measure. His grandfather wanted to write a book about decades of experience, but organizing the words and typing them were formidable obstacles. Yan, then a researcher who had spent years working in artificial intelligence, saw in that small family wish a reason to think bigger. A useful machine would have to meet people where they were, with whatever skills and tools they already had. A conversation could be easier than a keyboard. A product could matter more than an elegant demo.
That thought helps explain MiniMax, the Shanghai company Yan founded at the end of 2021. ChatGPT had not yet arrived. In China, betting on general-purpose AI was far from a comfortable consensus. Yan had a mathematics degree, a doctorate in AI and a senior job at SenseTime, where he had risen to vice president. He left to build models capable of handling text, voice and images, while also asking what ordinary people might do with them.
The founding idea sounds obvious only in retrospect. Train the model and build the application. Then discover, in public, what each one reveals about the other. It is a demanding way to run a young company. It requires research talent, product judgment and enough money to keep both moving. Yan still talks about those demands with the precision of a mathematician: resources, latency, error rates, cost per useful answer.
A book his grandfather could not write
Yan grew up in a relatively underdeveloped part of China, and he has spoken about how much people outside prosperous cities could gain from accessible AI. His path led through Southeast University, where he earned a mathematics bachelor’s degree in 2010, and the Institute of Automation at the Chinese Academy of Sciences, where he completed an AI doctorate in 2015. Postdoctoral research at Tsinghua University followed. At SenseTime, he worked for more than six years and became vice-head of its research institute as well as vice president.
The academic record is substantial: MiniMax’s corporate biography credits him with roughly 200 papers and more than 30,000 citations. Yet the family anecdote points toward a different measure of progress. Could a system help a person tell a story that might otherwise stay unwritten? Yan’s own formulation was that AI should become a product used in everyday life. He described the company’s aim as “Intelligence with everyone,” a phrase with an unusually practical test. If people cannot use the technology, the technology has missed its audience.
MiniMax’s first months did not deliver a neat product answer. Yan says the team spent six or seven months building a primitive model before products were possible. It hired a product manager early, even though he could not yet describe the eventual product. The ambition was an intelligent agent that could carry on a free, sustained conversation. The form was open. That combination of conviction and uncertainty would become a recurring feature of his decisions.
The avatar that looked wrong on a phone
An early version had a 3D character. Yan says the team filmed two models and placed the result on a handset. The effect was immediate: a figure staring continuously from a phone felt strange. They dropped the idea. The lesson had little to do with what a system could render and everything to do with what a person would want to live with. Voice, text and visual understanding remained central to the work, but the human-facing form needed to be lighter.
The first consumer experiments found an audience among AI enthusiasts and anime fans. Glow, an AI character chat product, launched in 2022. Rather than begin with elaborate A/B tests, Yan says the team read user feedback and watched how people used the product on social media, then checked its observations against data. Character conversations could look frivolous beside foundation-model research. For Yan, they were a demanding test of whether the model could respond well enough, fast enough and cheaply enough for repeated use.
MiniMax then widened the range: conversational products, audio, image and video generation, and tools for developers. Hailuo AI became its visual-generation platform. There was no single perfect format waiting to be found. Yan’s argument was that products and models would both improve through repeated attempts, even when an attempt was discarded. The abandoned avatar is memorable precisely because a founder who thinks in equations let an awkward feeling on a phone overturn the plan.
The expensive way to learn
The model work offered larger and costlier corrections. In 2023, MiniMax committed heavily to a mixture-of-experts architecture, which activates parts of a model for each task instead of running every part every time. The approach promised a way to build and serve larger systems under real compute limits. Yan said more than 80 percent of research resources and compute went into it. Two training attempts failed.
He described watching a training run drift away from its expected course over weeks, then tracing the problem through experiments, networks and data. The loss was money, but he emphasized time. Each failed run consumed weeks in a field where rivals were moving quickly. Looking back, he said the fixes were not confined to the architecture; they also made the research organization more rigorous. It is an unusually candid account of how a model team learns: by paying for mistakes that cannot be wished away.
Yan made another choice that sounds counterintuitive for a company hungry for computing power: he said MiniMax rented its GPUs rather than owning them. His concern was that a warehouse of owned hardware could change the business itself, making the company optimize for the assets it possessed. Renting had its own risks, including shortages. His answer was to become a large and dependable customer. The choice reveals his habit of asking how each resource changes the decisions around it.
He is sometimes described as quiet and attentive, a manager who listens before offering a pointed observation. Onstage at a 2024 conference, an investor recalled their first meeting in late 2021. Three people attended; only one left convinced enough to back the idea. The investor joked that the firm learned to bring three people to important founder meetings, in case one understood. The joke captures how odd MiniMax’s initial pitch sounded before the market had a familiar name for it.
A public test of a private conviction
MiniMax went public in Hong Kong on January 9, 2026. The listing made Yan’s early wager visible in a new way. The company now had consumer products, a developer platform, outside investors and quarterly numbers by which strangers could judge it. Yan, listed by MiniMax as founder, chairman, CEO and CTO, had to speak in two registers at once: research possibilities and business arithmetic.
The first half of 2026 shows both. Revenue reached $116.6 million, up from $30.4 million in the comparable period a year earlier. Open Platform and other enterprise services produced $73.9 million; AI-native products produced $42.6 million. Gross margin improved to 17.9 percent. The adjusted net loss, however, was $293 million. These are company figures, and they leave a clear question for Yan’s strategy: can the cost of delivering better intelligence fall quickly enough as usage rises?
MiniMax reported US$116.6 million total revenue for the six months ended June 30, 2026.
Yan put the tension plainly in August: “Intelligence can scale almost without limit; energy and compute cannot.” By July, he said, token consumption on MiniMax had grown to 20 times its January level. The number is impressive, but the sentence around it matters more. If AI is to reach beyond early adopters, each unit of useful work has to become cheaper. A voice conversation, a short film, a developer call and a long writing session all draw on the same underlying budget of computation.
That is why Yan’s public comments return so often to access. In a 2026 interview, he shrugged at the “AI tiger” label applied to Chinese startups and said the work was to improve models, revenue and service for a global audience. He described DeepSeek founder Liang Wenfeng as a friend while locating the larger competitive pressure among global technology giants. The observation is less theatrical than the animal metaphor and more revealing of how he sees the field: a long contest over capability, price and reach.
There is a pleasingly stubborn thread from his grandfather’s proposed book to MiniMax’s financial statements. The first is about a person who needs help expressing something. The second is about whether a company can make that help available at a sustainable cost. Between them sit thousands of engineering choices, several product experiments and some expensive failures. Yan’s story is still being written in that gap. He has made AI more conversational, visual and available through MiniMax. The remaining question is how many people can afford to make it ordinary.