This article is written based on public information and is for informational exchange purposes only, and does not constitute any investment advice.
On January 9, 2026, MiniMax (0100.HK) was officially listed on the Main Board of the Hong Kong Stock Exchange, becoming one of the first foundation model unicorns to complete its secondary market capitalization. MiniMax was listed at an issue price of HKD 165, surging 109% on its first day of trading, with its market capitalization briefly exceeding HKD 410 billion.
However, as of July 24, 2026, the stock price had reached HKD 196, with a market capitalization of only HKD 68.4 billion. Zhipu AI, which was listed during the same period, still maintains a high valuation of HKD 575.9 billion. The pricing given by the capital market differs by more than an order of magnitude.
Breaking down the financial report, we can see that MiniMax's fundamental performance is actually quite solid:
○ Total revenue in 2025 was RMB 553 million, a year-on-year increase of 158.9%, with over 70% of revenue coming from international markets, far exceeding market expectations.
○ Gross profit was RMB 141 million, a year-on-year increase of 437.2%. The gross margin was 25.4%, up 13.2 percentage points year-on-year.
○ Adjusted net loss was RMB 1.756 billion, compared to an adjusted loss of RMB 1.709 billion in the same period last year. The adjusted net loss margin narrowed significantly year-on-year.
(Note: The exchange rate is calculated at USD 1 = RMB 7, the same below. The financial report snapshot is as follows, in units of RMB 100 million, with all data sourced from Wind.)
Reviewing the financial report, we believe the essence of this divergence is that the market is evaluating foundation model enterprises with a new set of criteria: Whether foundation model enterprises can build a positive cycle of "foundation model capabilities—commercialization results—Infra capital expenditure."
Due to the phased weakness of M3, MiniMax is stuck at the first link of this cycle.
01The Foundation Model Company with the Broadest Business Model Coverage
MiniMax's "foundation model capabilities" performed exceptionally well prior to its IPO. Benefiting from the founder's spontaneous promotion after MaxClaw went viral at the beginning of the year, MiniMax's foundation model API call volume once ranked at the top of OpenRouter, enjoying immense popularity for a time.
More importantly, compared to foundation model enterprises listed during the same period, MiniMax has broader business model coverage. From the perspective of revenue structure, it faces the consumer market with Talkie and Xingye, the enterprise market with its open platform and enterprise services, and also has a product matrix including Hailuo AI. In 2025, MiniMax achieved a total revenue of RMB 553 million, a year-on-year growth of 158.9%, exceeding market expectations by 11%.
It is not only about full coverage of the consumer and enterprise markets at the product level, but also a broader sales market. In 2025, MiniMax's overseas revenue reached RMB 392 million, a year-on-year growth of 170.3%, accounting for approximately 73% of total revenue, up 3 percentage points year-on-year. The vast majority of overseas revenue came from the contribution of AI-native products, such as Talkie and Xingye.
According to the financial report, AI-native products contributed 67.2% of the company's revenue, while the open platform and enterprise services contributed 32.8%. The revenue structure shows a parallel growth trend between the consumer and enterprise markets, with the latter's growth rate (197.8%) significantly faster than the former's (143.4%).
Zhipu AI, another foundation model enterprise that listed on the Hong Kong Stock Exchange earlier this year, derives the vast majority of its revenue from localized deployment services for government and enterprise clients, with negligible contributions from consumer business. Therefore, the feedback from the capital market was also different in the early stage of listing. Zhipu AI's valuation anchor leans more towards AI infrastructure, and its stock price volatility at the beginning of the listing was significantly smaller than that of MiniMax.
The market's recognition of the direction of foundation model platforms developing consumer applications was relatively high at the beginning of the year, and MiniMax's stock price once reached a peak of HKD 410 billion. However, it didn't take long for the launch of Seedance to make the market realize that the barriers for consumer products are not that high.
Moreover, from MiniMax's perspective, the cost of acquiring users for AI Companion applications like Talkie in overseas markets is extremely high.
Breaking down the sales costs, as of September 30, 2025, MiniMax's sales and marketing expenses were approximately RMB 360 million, of which business promotion expenses accounted for as much as 92%. Advertising investment for consumer products is far higher than for enterprise products, and nearly 70% of MiniMax's revenue comes from overseas markets, with traffic acquisition costs priced in USD. Previously, Talkie's overseas advertising intensity was extremely high, and USD settlement further pushed up the absolute value of expenses. Although consumer products contributed about 67% of the revenue, the gross margin was only 4.7%.
Although marketing expenses fell in 2025, in the short to medium term, the high CAC (Customer Acquisition Cost) investment will not change, and consumer companion and entertainment products face the issue of a relatively high natural churn rate.
The commercial essence of the AI companion track lies in "monetizing emotional relationships." However, unlike real social networks such as WeChat and Discord, which are built on two-way interpersonal connections and have extremely high switching costs, AI companionship is one-way and highly dependent on algorithmic responses.
Under the impact of platform disclaimer clauses (such as data deletion limits), the emotional assets built by users over time are highly susceptible to turning into unguaranteed digital bubbles. Relationship assets lacking stability and persistence determine that companion apps cannot form solid switching barriers, which directly explains why they exhibit the traffic decay characteristics of "high attraction, low retention."
02Timely Correction, Significant Cost Control, and Improved Profitability
From the perspective of profitability, MiniMax is shifting from burning cash for growth to efficiency-driven growth.
In 2025, gross profit increased by 437.2% from RMB 26 million in 2024 to RMB 141 million in 2025. Its growth rate is significantly higher than the revenue growth rate. The gross margin increased from 12.2% in 2024 to 25.4% in 2025, doubling year-on-year.
In 2025, MiniMax's net profit attributable to shareholders was USD -1.913 billion, resulting in a net profit margin of approximately -356.24%.
However, after excluding non-operating items such as changes in the fair value of preferred shares, share-based payments, and listing expenses, the adjusted net loss in 2025 was approximately RMB 1.756 billion, a slight increase of only 2.8% compared to RMB -1.709 billion in 2024.
The improvement in profitability is mainly due to three aspects.
First, revenue from the enterprise open platform surged, with a significant increase in the payment conversion rate and retention rate of developers and enterprise clients (such as Agent developers, audio and video creation agencies). The proportion of high-gross-margin enterprise revenue in total revenue rose rapidly, structurally significantly boosting the overall comprehensive gross margin.
Second, it is due to the improvement in model and system efficiency, as well as the optimization of infrastructure configuration. Through MLA (Multi-head Latent Attention) and KV Cache quantization technologies, MiniMax squeezed VRAM (Video RAM) usage by over 50%. GPU utilization increased from the early 15%–20% to over 50%, and the inference cost per Token dropped exponentially, significantly improving the compute output ratio.
Third, and most importantly, MiniMax proactively reduced its investment in the consumer market. From the perspective of the full-year timeframe, sales and marketing expenses in 2025 fell by 22.3% year-on-year to RMB 366 million, and the sales expense ratio dropped to 68%.
03The Gap of M3: MiniMax Has Not Yet Crossed the Critical Threshold
In June 2026, MiniMax released its new generation foundation model, M3. Expectations were high before the release, after all, MiniMax had shown strong performance in foundation model capabilities, and its model call volume once ranked at the top of OpenRouter when MaxClaw went viral at the beginning of the year. However, the actual response after the release of M3 showed a clear gap with expectations.
According to data from the OpenRouter platform, the best result after the release of M3 was ranking 3rd in the second week, still failing to surpass the slightly earlier released DeepSeek V4 Flash. Moreover, due to pricing issues, the subsequent response was mediocre, and its current ranking has dropped to 8th.
The relative weakness in foundation model capabilities directly affected MiniMax, preventing a sensational effect at the product level similar to Kimi K3, and indirectly affecting the willingness of capital expenditure at the Infra level.
The reduction in Token inference costs is conducive to API sales, but as competitors like DeepSeek and Zhipu AI push Token prices to extremely low levels, the gross margin of the API business has been significantly compressed. Shortly after the release of M3, MiniMax announced a permanent 50% discount on API prices. JPMorgan Chase pointed out bluntly that this reflects "the model failed to gain market recognition for its original pricing and price increases," indicating that the cost advantage brought by algorithmic efficiency has not been translated into pricing power.
From a financial perspective, looking at the R&D expenses and capital expenditures disclosed by MiniMax, "low-cost model training" has not substantially reduced overall R&D investment. In 2025, MiniMax's R&D expenses reached RMB 1.781 billion, with the absolute amount increasing by 33.8% year-on-year.
Breaking it down, in the first nine months of 2025, cloud computing service expenses related to training in MiniMax's R&D expenditures were approximately RMB 987 million, accounting for 79%. Most of the costs were actually paid to Alibaba for compute supply, and its own assets did not show significant changes.
The asset structure also confirms this. In 2025, MiniMax's net fixed assets were approximately RMB 11 million, with production assets/total assets at 0.37%. For comparison, Zhipu AI's production assets/total assets were about 13.5%, making MiniMax only 1/36 of the former. Renting compute means having to submit to others' resource scheduling, which directly affects the autonomy of the model iteration pace. A deeper issue is that if foundation models cannot drive Infra capital expenditure, it essentially deviates from the underlying logic of the AI industry, which is asset-heavy and compute-heavy.
The leading advantage in foundation model capabilities is not clear enough, and the commercialization path for consumer products is full of uncertainties. The superposition of the two makes large-scale infrastructure investment a high-risk decision. The management's choice to wait and see is, fundamentally, a rational financial judgment. But it is precisely this wait-and-see approach that, in turn, limits the iteration speed of foundation model capabilities, naturally affecting the market's pricing logic for MiniMax.
04Foundation Model Enterprises Are Heading Towards K-Shaped Divergence
Up to this point, we have broken down the gains and losses of MiniMax across three dimensions: revenue structure, cost control, and foundation model iteration. However, behind the individual case of MiniMax is an industry-level trend that is accelerating: Foundation model enterprises are heading towards K-shaped divergence.
The logic of this divergence is not complicated.
In the foundation model track, foundation model capabilities are the starting point. Whoever has a stronger large model can obtain stronger feedback at the commercialization level: higher call volumes, more stable customer stickiness, and more substantial revenue. Once the positive feedback of commercialization is formed, enterprises can see clear directions for investment and dare to increase capital expenditure at the Infra level.
And the investment in compute and infrastructure will, in turn, accelerate the iteration of foundation model capabilities, forming a self-reinforcing upward curve.
Conversely, if foundation model capabilities are not prominent enough, the ceiling for commercialization will be pressed very low. Call volumes cannot go up, customer churn rates are high, and revenue growth is weak—these signals superimposed together make management hesitate to easily bet on large-scale capital expenditure. If compute infrastructure cannot keep up, the speed of foundation model iteration will be limited, and the gap will be further widened.
This divergence is simultaneously amplified in the capital market.
Players with strong foundation models who dare to invest will be regarded by the market as "core assets of large model infrastructure," gaining continuous capital inflows and valuation premiums. Meanwhile, enterprises with ambiguous foundation model capabilities and hesitant capital expenditure will be quickly abandoned by capital. After all, with the explosion of the open-source community (such as DeepSeek V4, K3), 85%-90% of the capabilities of LLMs (Large Language Models) and multimodal models have become public infrastructure for the entire industry. The technological generation gap has shortened from several years in the past to two or three months now, and the technological premium has rapidly dropped to zero.
MiniMax is at the crossroads of K-shaped divergence. MiniMax's most unique advantages are the breadth of its commercialization coverage, its execution capability in overseas markets, and its ability to optimize technical efficiency. These are all real capabilities accumulated by MiniMax over the past few years.
At the same time, the current difficulties are real. As it stands, the actual data of M3 is not as good as the iterative models of its peers, so its pricing has also faced some criticism, and the positive cycle at the capital level is gradually stagnating. More importantly, MiniMax still has the pressure of convertible bonds hanging over its head.
However, the endgame of the large model track has never been determined by a single release or a quarter's data. What truly determines the winner is whether they can cross the dividing line between "daring to invest only after seeing results" and "believing first to see." (Author: Xinghe)