In the spring of 2026, the winds shifted in the large-model industry. Capital is running out of patience, and the frenzy is ebbing away. Investors now care about only two things: whether your technology is truly irreplaceable, and whether the business model can actually scale into profitability.
Recently, the annual report of China’s first listed large-model company MiniMax has delivered its first report card since going public in January 2026. There is an important question than the data: And does it really deserve its current market cap of more than HK$250 billion?
Minimax has big-name shareholders, including tech giants Alibaba and Tenent.
The Cost of Growth
Let’s start with the fundamentals.
For full-year 2025, MiniMax posted revenue of US$79.038 million, up 158.9% year over year, with overseas revenue accounting for more than 70% of the total. But fourth-quarter revenue was about US$26 million, and the year-over-year growth rate fell from 175% in the first three quarters to 130%. The slowdown in the growth cadence suggests the growing pains of a strategic shift are starting to show.
Even more noteworthy is the change in the revenue mix.
Revenue tied to subscriptions and ongoing API usage rose from US$4.83 million to US$31.86 million—up 6.6x—and its share of total revenue jumped from 15.8% to 40.3%.
This metric has greater explanatory power than overall revenue growth. What it really signals is that customer stickiness is strengthening, and usage is shifting from occasional to habitual.
The loss figure does look alarming. Full-year net loss was $1.872 billion, but that includes $1.6 billion in accounting losses stemming from the conversion of convertible notes into equity. Excluding that factor, adjusted net loss was $251 million, up 2.7% year over year.
More importantly, sales expenses are shrinking: down from $87 million to $51.9 million, a drop of 40.3%, while revenue doubled over the same period. That drove the ratio of sales expenses to revenue down sharply from 285% to 65.7%. The old playbook of burning cash for growth is giving way to an efficiency-driven approach.
Gross margin rose from 12.2% to 25.4%, which looks solid. But when you pair that with the segment data disclosed in the prospectus, the picture is clearer: over the first three quarters, the consumer-side gross margin was only 4.7%, while the enterprise-side was as high as 69.4%. Add to that the fact that in Q4, the enterprise segment’s share of revenue increased from 29% to 41%, and the inference is straightforward: the improvement in overall gross margin was driven mainly by higher-volume, high-margin enterprise business, while the consumer segment’s profitability did not meaningfully improve.
To be fair, this is not necessarily a negative signal; rather, it’s a necessary strategic trade-off: consumer-facing products take on the job of accumulating interaction data and gathering user feedback, building a data flywheel that lays the groundwork for exporting capabilities to enterprise clients. The startup's founder Yan Junjie underscored this on the earnings call as well: the iteration of the dialogue model M2-her benefited from the massive volume of interactions on Xingye and Talkie, while the refinement of its video model was backed by the 600 million videos generated cumulatively by Hailuo AI.
That said, whether this model—“burn cash on the consumer side to buy data, and monetize on the enterprise side to fill the gap”—can form a sustainable closed loop still needs time to prove out.
Coding Capability: MiniMax’s Strength
Advances at the model level are the key pillar underpinning MiniMax’s bet on a “code-to-agent” closed loop.
In February 2026, it released M2.5, a foundation model positioned around an “agent-native design.” It scored 80.2% on SWE-Bench Verified, essentially on par with Claude Opus 4.6’s 80.8%. This performance directly answers the industry’s demand for agents’ core competency—coding ability.
Cost advantage is its other ace. M2.5 is priced at $0.30 per million tokens for input, and $1.10–$2.40 per million tokens for output. By comparison, Claude Opus 4.6 costs $5 for input and $25 for output—10 to 20 times more expensive.
Even if there’s still a capability gap, for users the real “punch” of this price-to-performance is that a continuously running Agent costs only about $1 per hour—and $10,000 is enough to keep four Agents running 24/7 for an entire year.
The surge in usage has validated the appeal of that cost-effectiveness. According to media reports, in February 2026, the M2-series text models’ average daily token consumption jumped more than sixfold from December 2025, while the Coding Plan package rose by over tenfold.
OpenRouter’s data is even more convincing: with monthly consumption of 2.45 trillion tokens, M2.5 rocketed to the top of the usage rankings, up 197% month over month, becoming the platform’s first Chinese model to surpass 50 billion daily tokens. Notion’s integration of M2.5 into its Custom Agents was also its first time choosing a non-European or U.S. model—an endorsement, to some extent, of M2.5’s technical capabilities.
But technological progress always comes at a cost. Benchmark results from Artificial Analysis showed that M2.5’s hallucination index fell from -30 for M2.1 to -41, with the hallucination rate climbing from 67% to 88%.
Behind this was a trade-off MiniMax made between model inference capability and the reliability of its outputs—because it wanted to seize a first-mover advantage in agents, it had to compromise on reliability.
But for agents built for enterprise scenarios, output reliability is precisely the core prerequisite. Once serious hallucination issues arise, they can undermine customer trust and even trigger business risks. This is also the key technical pain point MiniMax will need to address going forward.
Competitive Advantages
In hindsight, with homogenized competition in the large-model industry intensifying, MiniMax’s early strategic choices look exceptionally clear-headed.
In 2023, as domestic large-model startups crowded into benchmarking themselves against ChatGPT and sank into the quagmire of homogeneous competition, MiniMax chose a differentiated path. Yan Junjie put it this way at the time: “We can’t create unique value on that path.” Based on that judgment, MiniMax concentrated its limited resources on three core directions: two consumer products, Xingye/Talkie; the Hailuo AI video model; and its open platform.
Three key data points have preliminarily validated the feasibility of this differentiated approach:
An overseas revenue share of 73% shows that its overseas market布局 has gained a solid foothold and that it has strong international competitiveness; the open platform’s growth rate nearing 200% highlights the vitality and growth potential of its developer ecosystem; and revenue doubling while marketing expenses fell by 40% confirms that the model of “organic diffusion driven by model reputation and the developer ecosystem” works—also reflecting continued improvements in operating efficiency.
Yan Junjie’s redefinition of a “platform company” quietly reveals MiniMax’s strategic ambition: in the internet era, a platform’s core was serving as a traffic gateway, and the competition was about the ability to aggregate traffic; in the AI era, a platform’s core is to be the definer of the intelligence paradigm, and its value can be distilled into “intelligence density × token throughput”—intelligence density determines a model’s core competitiveness, while token throughput determines the scale and potential of commercial monetization.
This understanding also defines the depth of MiniMax’s strategy, allowing it to move beyond mere model-to-model competition and pivot toward competition over ecosystems and paradigms.
But a clear-headed strategic choice has not fully avoided the shared anxieties across the industry.
In late February, sentiment swings in Hong Kong-listed AI stocks unexpectedly exposed the fragility of the large-model sector: Zhipu saw its services come under strain due to a surge in traffic, and its share price fell 22% in a single day; although MiniMax did not suffer a direct service outage, it was still dragged down by sector sentiment, dropping 13%.
This volatility was not an isolated case; rather, it reflected a deeper, common predicament as AI startups move from the early stage into the growth stage—an “growth cliff”: when the user base and invocation volume rise rapidly, whether technical infrastructure, service capacity, and the profit model can keep pace in lockstep becomes the decisive test of whether a company can sustain its development.
An even deeper concern lies in path dependence in the technological roadmap.When users choose MiniMax, the core reason is often “close to top-tier models, but cheaper”; when they choose not to, the reason is just as blunt: “If the gap isn’t that big, why not use the original?” This latecomer’s “reference-frame dilemma” is like an invisible shackle— even if it pushes cost-performance to the extreme, it may still be forever chasing within a track defined by others, struggling to develop true irreplaceability. This is also a shared challenge faced by all late-starting large-model companies.
In addition, copyright risk hangs over the company like a sword of Damocles, ready at any moment to affect its growth. In September 2025, Hollywood studios including Disney sued MiniMax in California, alleging that its Hailuo AI engaged in infringement during model training, content generation, and promotion, with potential damages that could reach tens of millions of U.S. dollars.
The case is still at an early stage, but where it goes will not only determine MiniMax’s financial costs; it could also reshape large-model companies’ data-compliance strategies, directly impacting their business models and adding uncertainty to MiniMax’s overseas expansion.
MiniMax’s Test to Come in Six Months
Yan Junjie’s three forecasts for industry trends in 2026 sketch out where MiniMax plans to focus its efforts: programming will move into L4–L5 intelligence, making a leap from “tool” to “coworker-level” collaboration; office scenarios will replicate the rapid growth trajectory seen in programming; and multimodal creation will achieve “ready-to-deliver out of the box,” with breakthroughs in mid-to-long-form video generation enabling AI-generated content to deliver real practical value.
Taken together, Yan expects token consumption to surge by one to two orders of magnitude, and the company’s ARR is poised to enter the US$1 billion range. But realizing that vision depends on capability breakthroughs in the M3 foundation model and the Hailuo 3 video model—two products that are the key levers for MiniMax to close the “code-to-agent” loop and bridge the “growth gap.”
The pace of R&D spending also supports the view that the strategic center of gravity is shifting. In 2025, the company’s R&D expenses were US$253 million, up 33.8% year over year—far slower than revenue growth; however, fourth-quarter R&D spending for the single quarter was about US$72.47 million, roughly 20% higher than the average quarterly level across the first three quarters.
As M3 and Hailuo 3 enter a critical phase of training, the pressure on R&D spending in 2026 will only intensify. The roughly HK$4.82 billion raised in the IPO this January has provided some financial cushioning, but the ferocity of the industry’s compute arms race has not eased.
By the end of February, the company’s ARR had surpassed US$150 million, implying monthly revenue of about US$12.5 million—nearly double 2025’s average monthly level. But it’s important to stay clear-eyed: for a company whose annual revenue is still under US$80 million, the valuation has never been anchored to today’s results, but to the market’s expectations for how its future agent strategy will play out.
Back to the opening question: when “code-to-agents” becomes the core of a closed-loop business model, which side is MiniMax really on?
It has cards to play: B2B revenue is accelerating, it has already gained a solid foothold overseas, the share of recurring revenue has jumped, and marketing efficiency continues to improve; the “scissors gap” between M2.5’s performance and cost, and the organizational efficiency of having nearly 90% of internal tasks completed by AI, both show that it has achieved milestone progress in both technology and commercialization.
But the burden on its back is hardly light: the consumer side contributes two-thirds of revenue yet generates almost no gross profit; a rising hallucination rate is the price of technical trade-offs and could also become a hidden risk to market trust; copyright lawsuits remain unresolved; and the most fundamental question is this: when the core pillar of the business model is “value for money” rather than “irreplaceability,” how long can such a moat really withstand the impact of industry giants?
2026 is a turning-point year for the large-model industry—and a watershed moment for MiniMax. Every move it makes at this technological inflection point could be a matter of life and death.
Two quarters from now, with the release of the M3 model, deeper penetration into office scenarios, and the rollout of mid-to-long-form video capabilities, the market will be able to judge more clearly whether this 250 billion is truly a value anchor—or just another expectation that ends up being disproven. By then, MiniMax’s future trajectory, and its answer to that question, will gradually come into sharper focus.
Explore more exclusive insights at nextfin.ai.
