“Gentle,” “Buddhist,” “low-key” – these are the words frequently used to describe Yan Junjie, the 36-year-old founder and CEO of MiniMax.

Yet on January 9, 2026, this seemingly most “Buddhist” founder orchestrated one of the fiercest AI IPOs in history. MiniMax shares soared 109% on their Hong Kong Stock Exchange debut, closing at HKD 345 (~$45) per share and delivering a market capitalization exceeding HK$100 billion (~$12.8 billion). It was the only Hong Kong tech IPO in the past four years to more than double on opening day, with 420,000 subscribers generating an oversubscription rate of 1,838 times. Another interesting detail is that MiniMax is China’s second “AI tiger” to go public – just one day after Zhipu AI.

Welcome yo our GenAI Unicorn series! We’ve covered other three AI Tigers before (Zhipu, Moonshot, Baichuan). Now it’s time for MiniMax. There are many striking details about this company. The decision to bet on multimodality from day one – back in 2021, long before it was fashionable. The late but decisive turn toward open source. The explicit rejection of the “genius founder” myth. An organization where researchers and engineers sit side by side, working simultaneously on foundation models and real products. And, perhaps most surprisingly, MiniMax appears to have solved what most AI startups treat as an impossible triangle: frontier-level performance, radically lower costs, and genuine mass adoption.

At first glance, the IPO looks like proof of success. But is it?

In China’s AI market, an IPO is not an endpoint. It is a way to buy oxygen. Companies go public early, often while still burning money, and are judged in real time rather than in the comfort of long private runways.

This is what makes MiniMax worth examining now. Their story – a high-stake experiment – just begins. How did they get here? Will they survive? Let’s discuss.

In today’s episode:

In December 2021, Shanghai’s tech scene was buzzing with anticipation for SenseTime’s imminent Hong Kong IPO – but a small group of its computer vision veterans were already plotting their exit. Veterans who were barely in their early thirties.

Yan Junjie had spent more than six years at SenseTime, rising to become one of its youngest vice presidents. He had seen the company at its strongest, scaling computer vision into real businesses, real contracts, real revenue. He had also seen its limits. Vision models were becoming projects, then products, then maintenance work. General intelligence was not part of the plan.

The conviction that another path was possible had been forming much earlier. In 2014, during a summer internship at Baidu, Yan worked on large-scale speech recognition and encountered GPU clusters for the first time. What stood out was not architectural cleverness, but regularity. As data and compute increased, performance improved in smooth, predictable ways. Dario Amodei – who would later co-found Anthropic – was at Baidu around the same time, working on similar problems. The pattern was visible to anyone paying attention: throw more GPUs and more training data at the problem, and accuracy went up in a smooth, almost boring curve.

"Scaling Law was discovered around 2014 when we were doing speech recognition," Yan would recall years later in his interview with Luo Yonghao. "It happened at a Chinese company. But then... nothing. We didn't capitalize on it.”

That regret would haunt him. China had seen the future first – and let it slip away.

The second revelation came in 2019, when Yan – a gamer himself, even known by his colleagues by his Dota nickname IO – was entranced by how OpenAI’s Dota 2 system defeated the world’s best professional teams. This was no longer narrow optimization. It was coordination, adaptation, and strategy under uncertainty. If reinforcement learning could scale there, Yan began to wonder where else it could scale.

And then OpenAI published “Language Models are Few-Shot Learners,” introducing GPT-3. That paper blew Yan’s mind. There were so many things possible with foundation models. Inside China’s AI circles, though, the idea was met with disbelief. Multimodal AGI – what the heck was that?

MiniMax registered in Shanghai in December 2021 with a small founding team drawn from the same ecosystem around SenseTime. Yan Junjie as CEO. Yun Yeyi, then director of innovative business at SenseTime, brought operational discipline and an outward-looking, global orientation shaped by four years in the US (she attended Johns Hopkins and Columbia Universities). A few articles list Yang Bin and Zhou Yucong as co-founders, but we couldn’t confirm that they are still involved in the company’s business. The MiniMax’s website lists Yan Junjie as the only founder. Yun Yeyi’s Linkedin shows her as a co-founder.

In any case, what united the team was a shared conviction. Intelligence could be treated as an engineering problem, broken down, optimized, and built under constraint.

The experiment had begun. And it needed money.

The Genshin Impact Gamble

Having a vision and getting funded are different games. The team’s first pitch meetings were brutal. Yan Junjie later recalled this period in one word: “painful.” In early investor meetings, their ambitious claims about building consumer AI products from scratch and working toward AGI met skepticism, sometimes outright dismissal. No one could imagine how powerful AI would become. Some investors in big tech circles would simply call him a fraud.

But a gamer sees another gamer. MiHoYo, the gaming studio behind Genshin Impact, co-founded by billionaire Cai Haoyu, decided to bet on Yan’s vision. In 2022, they recognized that the next generation of gaming experiences would be AI-native, with dynamic NPCs and procedurally generated content requiring exactly the multimodal models MiniMax proposed to build.

With gaming industry credibility in hand, the MiniMax team secured their first formal round: approximately $30 million in early 2022 from Yunqi Capital, where Chen Yu (now managing partner) became their first institutional believer. The bet was counterintuitive: while most Chinese AI startups chased B2B enterprise contracts, MiniMax declared from day one they were "B2C first."

That was bold.

AGI Through Consumer Products

From the beginning, MiniMax’s stated mission has been unambiguous: “Intelligence with Everyone” – not intelligence for everyone, but with everyone, positioning AI as a collaborative partner rather than a tool. Yan would later say that one of the inspirations for him was his grandpa, who wanted to write a book. Yan believed that AI would be a perfect partner for him. That’s why he needed to build it. Their path to AGI differs fundamentally from peers. While Zhipu AI pursues alignment with human values through academic research and enterprise solutions, MiniMax believes the route to AGI runs through consumer interaction at massive scale.

Surprisingly enough, with that goal, he was not focusing on metrics. Yan Junjie insisted that the mobile internet logic of “more users equals better products” doesn’t directly apply to AI development. “We’re not optimizing for MAUs [monthly active users],” he emphasized. “We’re optimizing for model capabilities that enable genuine intelligence.” This might sound like typical startup rhetoric, except MiniMax actually built their entire technical stack around this principle.

"AGI must serve ordinary people, including the elderly and children," he says. "True general intelligence requires multimodal inputs and outputs – language, images, audio, video. Consumer applications generate diverse, messy, creative data necessary to train truly general intelligence. Enterprise use cases optimize models for narrow, predictable tasks." That was another thing that made MiniMax special: they were workin on multimodal model from day 1, way before it became a thing.

The "Anti-Genius" Philosophy

What distinguishes Yan's approach is his rejection of hero worship. "AI is not mysticism. It is an engineering problem that can be decomposed using first principles," he states flatly. "Algorithm design, data pipelines, training efficiency – every link has a clear objective."

This philosophy extends to hiring. Unlike Silicon Valley's cult of recruiting "10x engineers," Yan spends enormous time finding like-minded individuals with excellent basic qualities. "I'm not a genius. I'm a winner of systems," he says. "Organizations and individuals can grow together."

Funny fact: a lot of MiniMax employees don't use real names internally – going by nicknames, including Yan, who chose "IO" after a support hero in Dota 2. When a reporter at the 2024 World AI Conference mentioned an employee's real name, colleagues struggled to match it to the person. This egalitarian structure reflects Yan's belief that environments matter more than individuals, and organizations can give birth to innovation.

AI's Three-Stage Evolution

Yan articulates AI's development in three clear stages:

Stage 1 (Before 2022): AI as a component embedded within products – facial recognition in security systems, speech recognition in voice assistants. Value was limited and fragmented.

Stage 2 (2023-2026): AI with general capabilities driving standalone products – ChatGPT, Midjourney, Character AI. Models have sufficient capability to be useful, but error rates remain 20-30% in evaluations and real scenarios.

Stage 3 (Now and the Future): AI performance stably surpasses average individuals. Error rates drop by an order of magnitude (to ~3%). User interaction time with AI exceeds time spent on traditional recommendation-based apps. This is the true inflection point requiring a tenfold error reduction and a hundredfold increase in application scale.

From the outside, MiniMax seemed to appear suddenly in 2023 with consumer apps and large models. Internally, the company had already spent close to two years building toward that moment.

Between late 2021 and early 2023, MiniMax did not focus on products at all. The team was testing whether the core assumptions behind frontier AI held under Chinese constraints. This meant training early general-purpose language models, validating scaling behavior, and running parallel experiments across text, speech, and early vision–language alignment. None of this was public. The goal was not visibility, but confirmation that scaling laws were real, reproducible, and exploitable without US-level compute budgets.

During this period, MiniMax also made its first attempt at Mixture-of-Experts (MoE) architectures. These attempts failed repeatedly. Training instability, routing inefficiencies, and inference costs initially outweighed any gains. The work was scrapped and restarted multiple times before the team reached something usable in late 2023. This period consumed a large share of compute and produced little external validation, which reinforced the perception that MiniMax was drifting without a plan.

On the model side, MiniMax first released the ABAB series, culminating in ABAB 6.5 in April 2024. These were among the earliest MoE-based language models deployed at commercial scale by a Chinese startup. The choice of MoE was driven by economics, not aesthetics. Dense models hit clear cost ceilings under sustained inference load. MoE was adopted to reduce per-token compute while maintaining model capacity. MiniMax has stated that ABAB 6.5 achieved performance close to leading U.S. models on selected benchmarks at substantially lower inference cost, though these comparisons were not independently audited.

In early 2025, MiniMax consolidated its stack under the MiniMax-01 family, comprising a general-purpose text model and a vision-language model. This marked the company’s first open-sourced release, signaling a shift toward faster external feedback and ecosystem trust. These models powered the expansion of Hailuo AI, a multimodal creation platform that added music, speech, and later video. The video models, first released in late 2024 and iterated through 2025, emphasized short-form generation with controllable motion and temporal coherence, and were deployed directly to consumers rather than positioned as research artifacts.

Audio followed the same pattern. Speech-02, released in April 2025, supported multilingual text-to-speech, low-latency interaction, and short-sample voice cloning. It became the default voice layer across MiniMax’s conversational products.

In mid-2025, MiniMax released M1, an open-weight reasoning model designed for long-context workloads. It supported context lengths of up to roughly one million tokens using hybrid attention mechanisms and internal training optimizations. A successor model, M2, released later in the year, focused on general-purpose and agent-oriented tasks and saw significant uptake through developer platforms.

In 2023, they also started releasing apps. A few highlights:

Talkie: The Virtual Companion Phenomenon

Launched in 2023, Talkie is an AI character chat application that allows users to interact with AI-generated personas – from celebrities (like Kim Kardashian) to fictional characters to customized companions. The app gained decent traction globally:

Unlike productivity tools, Talkie fulfills social and entertainment needs, generating massive engagement that feeds back into model improvement. However, the app has faced regulatory scrutiny. In late 2024, it was temporarily removed from the U.S. App Store amid concerns about Chinese-owned AI applications. By December 2025, Chinese regulators also announced stricter oversight of AI chatbots to prevent content related to suicide, gambling, and other sensitive topics, potentially affecting Talkie's operations domestically.

Hailuo AI: Hollywood Studio in Your Pocket

Launched in March 2024 as a multimodal consumer platform, Hailuo AI began with text and music generation before adding its viral video feature in September 2024:

Their success attracted legal attention: in September 2025, Disney and Universal filed copyright lawsuits alleging that MiniMax trained Hailuo AI on copyrighted material featuring their characters without permission.

MiniMax did not begin as an open-source company. For its first two years, the models remained closed, iteration happened internally, and there was little public signaling. As Yan later reflected, the team underestimated how important technical trust and technical reputation are in an industry where progress spreads through shared understanding as much as through raw compute.

The change came gradually in 2024 and 2025, and it reflected a shift in how MiniMax understood defensibility. Model quality, Yan observed, converges far more quickly than ecosystems do. As more architectures, training methods, and evaluation techniques become public, sustained differentiation at the model level becomes increasingly difficult.

From this perspective, openness becomes a strategic necessity. Yan has argued that companies outside the clear first position have little choice but to open up if they want to remain relevant. He often uses the comparison between Android and iOS: Apple can afford to remain closed because it controls the end-user relationship, while others rely on openness to build adoption, trust, and collaboration.

When MiniMax open-sourced the MiniMax-01 family and later the M1 and M2 models, it did not abandon commercial discipline. Model weights were released, but training infrastructure, large-scale deployment, and product integration stayed internal. Academic transparency was not their goal. They aimed at faster feedback cycles, stronger external scrutiny, and a broader base of contributors.

This approach also aligns with a deeper belief that runs through Yan’s thinking. User scale does not directly translate into smarter models, and DAU is a weak proxy for intelligence. If improvement depends primarily on research iteration rather than usage volume, then openness becomes more valuable than reach. In that sense, MiniMax’s turn toward open source reflects a sober assessment of how advantage is built and lost in frontier AI.

MiniMax does not make money in the way most Chinese AI startups were designed to.

From the beginning, the company avoided a project-driven enterprise model built around customized deployments, government contracts, or one-off integrations. That path offers early revenue, but it locks models into narrow optimization targets and fragments the technical stack. Yan Junjie has been explicit that this was a trade-off MiniMax chose not to make.

Instead, MiniMax operates a dual revenue structure, with both sides built on the same underlying models.

Yan has repeatedly argued that over-optimizing for revenue too early distorts technical decisions and slows long-term progress.

MiniMax’s monetization strategy reflects a longer view. Products generate revenue, but also usage patterns. Enterprise APIs generate cash, but do not dictate model direction. Open source accelerates learning rather than cannibalizing income.

In short, MiniMax makes money by serving users at scale with general-purpose models, while delaying the point at which revenue becomes the dominant constraint on technical choices. Whether this balance can hold as a public company remains an open question, but the structure itself is intentional rather than accidental.

Founded in late 2021, MiniMax completed seven funding rounds in just over four years, with its valuation rising from USD 170 million at the angel stage to USD 4.2 billion shortly before IPO. On January 9, 2026, MiniMax successfully listed on the Hong Kong Stock Exchange. On its first trading day, the stock surged 109%, pushing market capitalization beyond HKD 90 billion (USD 11.5 billion), making MiniMax one of the fastest companies to reach IPO in China’s AI sector.

MiniMax occupies a unique position among China's "Four AI Tigers" (sometimes called "Six Little Dragons" when including two additional startups):

Zhipu AI: Academic pedigree from Tsinghua University, focuses on alignment research and enterprise solutions. Listed one day before MiniMax with more modest first-day gains (~30% vs. 109%).

Moonshot AI (Kimi): Founded by former Google Brain researcher Yang Zhi, known for extremely long context windows (200k+ tokens before MiniMax-01) and enterprise focus. Reportedly declined IPO to pursue additional private funding.

Baichuan Intelligence: Founded by former Sogou CEO Wang Xiaochuan, targets enterprise vertical solutions. More B2B-oriented than MiniMax.

The differentiation is clear: Zhipu has academic prestige, Moonshot has enterprise contracts, Baichuan has search-engine heritage – but MiniMax has consumer traction at scale. This is both advantage (defensible user base, diverse training data) and risk (regulatory exposure, platform dependency).

Several factors distinguish MiniMax from both Chinese and international peers:

Despite the spectacular IPO, MiniMax faces significant headwinds:

Similar lawsuits have targeted Stability AI, Midjourney, and others – this is an industry-wide issue, but MiniMax's consumer focus makes them a higher-visibility target.

In Yan Junjie’s view, the China-US AI gap is real, but often misunderstood. American frontier labs operate with orders of magnitude more capital, more stable access to advanced chips, and longer private runways. Chinese startups operate under permanent constraint. Everyone has ambiton – what differs is the optimization problem.

Yan has repeatedly argued that US companies can afford inefficiency. When compute, capital, and talent are abundant, brute-force scaling works, at least for a while. In China, the same approach collapses quickly. That forces a different path, one centered on architectural efficiency, cost control, and tighter integration between models and products.

This difference shows up in priorities. US labs tend to push single-modality frontiers first, then layer products later. Chinese companies are pushed to think about deployment, inference cost, and user accessibility much earlier. According to Yan, this is not a disadvantage by default. It creates pressure to innovate where returns compound fastest.

Yan does not believe AI will converge to a single winner. He expects a small number of global players, with at least one or two coming from China, each shaped by different trade-offs. In this framing, the competition is not about copying OpenAI’s trajectory, but about discovering alternative ways to reach similar capability ceilings under different constraints. Copying OpenAI’s trajectory might not be a good idea at all – they also still need to prove that they can be profitable.

The more important divergence, in his view, is philosophical. The mobile internet rewarded traffic accumulation and feedback loops. AI does not follow the same logic. More users do not automatically produce better models, and excessive scale can even slow iteration. This is where Yan believes many observers misread the China–US comparison, focusing on surface metrics rather than underlying dynamics.

From MiniMax’s perspective, the future of AI is not decided by who scales fastest in absolute terms, but by who learns fastest under constraint. The United States currently leads on raw capability. China is being forced to explore efficiency, multimodality, and deployment-first thinking earlier. Whether that path produces durable advantages remains unresolved, but it is no longer peripheral to the global AI story.

MiniMax’s journey from rejected pitches to an $11.5 billion public company in just four years is a case study in strategic contrarianism. Now, MiniMax is also an increasingly important experiment: whether AI economics can work outside infinite VC subsidy.

Every frontier lab operates on faith that capability converts to profit eventually, somehow. OpenAI loses billions while searching for sustainable business models. Anthropic has raised many billions of dollars while revenue remains small relative to its capital base. Much of the industry is built on deferred monetization, with the assumption that frontier performance will justify today’s burn tomorrow.

MiniMax is testing a different hypothesis under public scrutiny. Its IPO demonstrates that a consumer-first, multimodal AI company built under persistent constraint can reach global scale, attract real users, and convince markets to fund continued experimentation. It does not demonstrate durability, profitability, or long-term defensibility, but it does make the alternative visible.

What MiniMax is running is a coherent system rather than a set of tactics. Intelligence is treated as something that improves through interaction rather than enterprise deployment. Multimodality is treated as foundational. Openness is used to accelerate learning without surrendering commercial control. Cost pressure forces architectural efficiency rather than being postponed by capital abundance. Consumer products generate diverse signals, open models invite scrutiny and faster iteration, and tight compute budgets shape design choices early.

The risks remain structural. Public markets compress timelines, consumer AI draws regulatory and legal attention first, and competitors with deeper pockets continue pushing raw capability forward. The IPO is not an endpoint – it is permission to keep running the experiment in public. Whether learning under constraint compounds faster than capital-driven scaling remains unresolved, but MiniMax has made the question concrete, measurable, and unavoidable.

Thank you for reading, please feel free to share with your friends and colleagues. 🤍