MiniMax CEO Yan Junjie Bets That More Users Will Not Build Better AI

Yan Junjie sat across from podcaster Luo Yonghao for four hours last fall, before MiniMax's Hong Kong IPO, and said something that has been quietly reshaping how the company operates: more users do not make your AI model better.

That claim cuts directly against the foundational logic of the internet era. The whole promise of recommendation engines, social platforms, and search was that user data compounds — the more people interact with a system, the smarter it gets. Network effects are the moat. Scale is the strategy.

Yan doesn't think that applies to foundation models. "It doesn't depend on how many users you have," he said in the interview, "but on quality, and whether you can find the distributions that truly make the model smarter." A recommendation engine trained on a billion low-quality clicks gets worse, not better. An LLM trained on well-curated, reasoning-dense data can improve with a fraction of that volume.

Whether that's a genuine theoretical insight or a convenient rationalization for a capital-constrained startup is the most interesting question the MiniMax story raises. The company's first post-IPO financial report, released March 2, offers some evidence for the former.

MiniMax generated $79 million in 2025 revenue, up 159% year-over-year. Gross margin improved from 12.2% to 25.4%. Sales and marketing spending fell 40%. That is not a company burning cash to acquire users. It's a company that bet the user flywheel didn't matter — and appears, at least through one financial cycle, to have been right.

The IPO itself was not modest. MiniMax listed in Hong Kong on January 9, raising $614 million. Shares doubled on debut to a $13.7 billion market cap — per Reuters, which reported closing shares at HK$345 versus the offer price of HK$165 — on an oversubscription of roughly 1,848 times, according to Caixin Global. Yan Junjie personally netted a $1.6 billion paper fortune, according to Bloomberg. The company had been burning roughly two billion yuan per month in the lead-up.

But Yan's real argument in the Luo interview isn't just about unit economics. It's about what kind of company MiniMax is. He frames the AI industry not as an extension of mobile internet but as a genuinely new organism — one that requires a different product philosophy, a different organizational logic, and a different definition of talent.

On talent: MiniMax and DeepSeek share an unusual profile. Both companies skew young, domestic, and internally grown. Yan started the company in 2022, before GPT-3.5, when the AI talent ecosystem was thin. The people he could recruit were mostly those who hadn't yet made names for themselves in traditional AI. He calls them "grassroots." His most valued hires now are often people for whom MiniMax is their first job. He tests not for pedigree but for the ability to collaborate, and for what he calls a belief in "first principles."

On organization: MiniMax tried OKRs and found them unworkable. You can't set a quarterly target for whether a model will converge. The company remains three layers deep — Yan, his direct reports, and their direct reports, across 400 people. He made the decision, apparently over significant internal resistance, to fuse the algorithms and infrastructure teams into a single unit optimizing a single objective rather than parallel scopes. Some people left over it.

On product: Yan describes traditional apps as "channels" in the AI era. The actual product is the model. MiniMax's consumer products — Talkie (international), Xingye (domestic), Haicao (video) — are showcases for model capability, not the source of it. DAU is what he calls a "vanity metric." That's a sharp stance for a company that built its early reputation on consumer apps, and it helps explain the marketing cut.

MiniMax-M1, released in June 2025, is the clearest technical expression of that philosophy. It is the world's first open-weight hybrid-attention reasoning model, combining a Mixture-of-Experts architecture with a Lightning Attention mechanism that enables linear computational complexity over long contexts. The practical result: M1 supports a 1 million token context window — 8 times DeepSeek R1 — and when reasoning at depth over 80,000 tokens, requires roughly 30% of DeepSeek R1's compute. On TAU-bench, the agentic tool-use benchmark, M1 outperforms all open-weight models and beats Gemini 2.5 Pro. On long-context understanding, it ranks second globally, trailing only Gemini 2.5 Pro.

The training cost for the entire RL phase was $534,700 — 512 H800s for three weeks. That number, if accurate, is remarkable. The company's novel RL algorithm, CISPO, clips importance sampling weights rather than relying on token-level updates, and converges at twice the rate of ByteDance's DAPO and significantly faster than DeepSeek's GRPO. MiniMax says the final training exceeded their own expectations.

The financial figures invite skepticism in a few places. The IFRS net loss for 2025 was $1.87 billion — up from $465 million the year prior — driven largely by non-cash fair value changes in financial instruments, a standard accounting artifact for recently-listed companies with volatile valuations. The adjusted operating loss was $250 million. That's still a company spending vastly more than it earns, and the $1.05 billion in cash reserves (pre-IPO proceeds) is not infinite runway at prior burn rates.

The early 2026 data is more striking. By February, daily token consumption for M2 text models had grown sixfold compared to December. Developer registrations quadrupled. Annualized recurring revenue crossed $150 million. If that trajectory continues, the margin story gets more interesting.

Yan's 2014 moment is worth sitting with. He interned at Baidu that year — the same company where a young Dario Amodei was observing scaling behavior in speech recognition that would seed the ideas behind Anthropic's founding. Two researchers, on opposite sides of the Pacific, watching the same phenomenon from different vantage points. Amodei left for OpenAI and eventually founded Anthropic around alignment. Yan left for SenseTime and eventually founded MiniMax around multimodal AGI. The parallel isn't exact, but it rhymes.

The MiniMax thesis, reduced to its core: intelligence does not compound from volume, it compounds from quality. The internet era taught us to maximize for scale because scale built flywheels. The AI era may operate on different physics. If Yan is right, the companies optimizing for user acquisition are not just wasting money — they're misreading the map.

Whether he's right won't be clear for years. But the fact that the company's first public financials support the claim rather than undercut it makes this more than a founder's contrarian pose. It's a testable bet, and for now, the early numbers are on his side.