MiniMax's stock price plummeted by over 72% in the six months since its Hong Kong listing, falling from a high of HK$410 billion to approximately HK$100 billion.Article author and source: Dingjiao
On July 9, nearly half of MiniMax's shares were unlocked.
On January 9, six months ago, it listed on the Hong Kong Stock Exchange, one day after Zhipu. On its first trading day, MiniMax surged 109%, far outperforming Zhipu and briefly becoming the more sought-after large model stock. In March, its share price peaked at HK$1,330, with a market capitalization exceeding HK$410 billion, briefly surpassing Baidu.
The stock price then reversed course. On June 1, the day the flagship model M3 was released, it opened higher but plunged sharply, closing down 15%; over the next two weeks, the stock price halved. By early July, it had fallen 72% from its peak. As of the close on July 3, its stock price stood at HK$346, with a market capitalization of HK$108.7 billion.
Zhipu later surged ahead, with its market capitalization once exceeding one trillion Hong Kong dollars, and now hovering around eight hundred billion Hong Kong dollars.
On July 8 and 9, the two companies each faced a six-month lock-up expiration, but for Zhipu, it was merely a minor test, while for MiniMax, it was a major challenge. Zhipu’s lock-up release amounted to only about 5.76%, whereas MiniMax saw approximately 44.85% of its shares unlocked, instantly expanding its potential float to around 50%.
Some investors following the Hong Kong stock market commented, "A short-term decline is normal—the previous stock price was inflated by too much sentiment and scarcity premium." Another industry insider emphasized to 'Dingjiao One': "Although it’s still up 110% from the offering price, investors’ sentiment has completely shifted."
Over the past six months, MiniMax's presence in China's large model market has diminished. In terms of model capabilities, its position in the top tier still needs strengthening; in terms of market visibility, it has not garnered as much attention as DeepSeek or Zhipu.
Liu Yujia, Investment Partner at Zhengjing Capital, believes that the recent price divergence between MiniMax and Zhipu over the past six months cannot be simply interpreted as one company having a stronger model; rather, the market is re-pricing different AI pathways. MiniMax’s strengths lie in multimodality, consumer-facing products, and overseas growth, but the market is still observing whether it can translate these advantages into agents, coding capabilities, a developer ecosystem, and stable revenue streams. Zhipu, on the other hand, aligns more closely with narratives around domestic foundational models, enterprise services, developer platforms, and China’s AI infrastructure, making it more likely to command a valuation premium under current market conditions.
However, stock performance is short-term. In his view, MiniMax’s real test is whether its models, products, and commercialization can achieve a closed loop. “If they can, the ceiling will be much higher. Otherwise, it may remain stuck in a ‘broad but shallow’ state.”
01 This round of decline has eliminated two layers of premium.
Although MiniMax's stock price has dropped significantly, the capital market has not entirely rejected this large model company. “It’s just that the market premium brought by the rally in March this year is gradually returning to rationality,” commented Leo, an industry professional focused on large model companies.
The decline is primarily due to premiums on two fronts.
First is the premium for model capability.
MiniMax positions itself as one of only four global leaders in multimodal AI, alongside OpenAI, Google, and ByteDance. Achieving this top-tier status implies control over pricing power and justifies higher market valuation multiples. However, after six months of market validation, this label warrants scrutiny.
The official claim states that M3 achieved 59% on SWE-Bench Pro, surpassing GPT-5.5; however, independent evaluation by Artificial Analysis’s Intelligence Index ranked it ninth among mainstream models, and on Chatbot Arena, which prioritizes real user preferences, it dropped to beyond the 40th to 50th positions. Multiple respondents noted that China’s top-tier large models are Zhipu, Kimi, and DeepSeek, with MiniMax’s performance being questionable.
Image source / MiniMax
More critically, one week after M3's launch, MiniMax permanently halved its API pricing. JPMorgan subsequently downgraded its rating from "Overweight" to "Neutral" and reduced its target price from HK$1,100 to HK$400, citing that "since the M2 model, MiniMax has not launched a new domestic SOTA (state-of-the-art) model and, purely in terms of model capability, it remains in a catching-up phase."
Zhipu, during the same period, raised its API pricing multiple times, with a cumulative increase of over 80%, while its usage volume grew fourfold.
A price cut caused such a significant market shakeup that Li Zeming, Chief Investment Officer of Hongyi Capital, said that Hong Kong stocks for large model companies are primarily valued using the price-to-sales ratio (market capitalization/revenue). “So whenever you hear that a major large model company has significantly lowered its prices, the entire sector tends to plunge simultaneously.” MiniMax’s price reduction lowered revenue expectations, thereby reducing its valuation.
Second is the premium from the C-end story.
In 2025, 67% of MiniMax's revenue came from consumer-facing AI products, primarily overseas emotional companion apps like Talkie and domestic ones like Xingye, along with the video tool Hai Luo AI. At the time of its listing, this was the fastest-growing business, and the market's high valuation largely bet on "consumer globalization."
However, C-end data significantly declined later on.
Monthly active users for Talkie and Hoshino declined by 60% quarter-over-quarter in Q4 2025. The emotional companionship sector as a whole is also tightening: Talkie was temporarily removed from app stores in certain overseas regions, and Hoshino in China underwent regulatory rectification.
The video segment is also under pressure; Hai Luo AI has lost its leading position on the Artificial Analysis video model leaderboard, surpassed by multiple models including Alibaba’s HappyHorse, ByteDance’s Seedance, and Kuaishou’s KeLing.
During this period, market perceptions of MiniMax’s model also shifted. Liu Yujia summarized it as a “relatively traditional but comprehensive” internationalization strategy, “somewhat similar to ByteDance’s early overseas expansion approach.” Initially, the market embraced this model, allowing MiniMax to benefit from valuation premiums in overseas AI applications; later, the rise of Agent ecosystems like OpenClaw created temporary opportunities for its models and API usage. However, as the industry narrative further shifted toward Agents, developer ecosystems, and toolchain platforms, the market began to reevaluate: Is MiniMax fundamentally a model company with several global products, or can it become the foundational model gateway of the Agent era?
02
It's easy to be big and comprehensive, but hard to be big and strong.
MiniMax's business model differs from that of other large model companies.
Among China’s large models, DeepSeek iterates its model based on community feedback, Zhipu relies on enterprise data from B-end clients, and MiniMax follows the most resource-intensive path. Its foundation is a proprietary, fully multimodal large model—developed in-house for language, video, audio, and music.
On this model, it runs two businesses simultaneously: for individual users, it launched a suite of AI-native apps focused on overseas markets; for enterprises and developers, it built an open platform that sells model capabilities via API based on usage volume. “In Leo’s view, this shift led to rapid revenue growth but also caused the company’s revenue structure to become overly dependent on the C-end.”
MiniMax’s strategy is based on the logic that models create products, which attract users and data globally, and this data continuously feeds back to improve the models. Liu Yujia believes that the three elements—model, product, and data—form a positive feedback loop, with an upper limit of “large and powerful” and a lower limit of “large and comprehensive.”
Currently, issues have arisen at the levels of the model, product, and commercialization.
The model layer has strengths but lacks uniqueness.
MiniMax has advantages in vertical capabilities: its speech synthesis model, Speech-02, once topped the Artificial Analysis global TTS ranking, and its text model, the MiniMax-01 series, is known for its ultra-long context window of 4 million tokens. However, these technical specifications have not translated into user engagement.
Li Zeming believes the technological gap between China’s top large models may be only one to two months: “If you’re ranked high this month, next month the entire ranking could have changed.” In an environment characterized by rapid technological iteration and unstable rankings, “irreplaceability” cannot be built solely on a single technological advantage—it requires sustained engineering leadership combined with user loyalty accumulated over time. Based on market feedback, MiniMax has not yet entered the top tier in terms of model awareness or user perception.
Product layer: four initiatives progressing in parallel, but the data pipeline requires validation.
MiniMax's public narrative is "fully self-developed multimodal": four data streams mutually feed and enhance each other. However, user behaviors across product lines vary significantly—Hailuo AI focuses on video creation, while Talkie and Xingye emphasize AI role-playing and emotional companionship. Talkie’s data can reinforce the emotional companionship model itself, but offers limited value in improving core capabilities such as programming or reasoning. Whether data can effectively flow across products and modalities remains uncertain.
Liu Yujia outlined three specific evaluation criteria: first, whether cross-modal transfer can be achieved—for example, can video data enhance the model’s understanding of the physical world and temporal relationships, and can audio data improve its judgment of emotion and rhythm? Second, whether disparate training frameworks for text, audio, and video can be consolidated into a unified engineering foundation model. Third, whether the overall capability is built on sufficiently strong individual capabilities, rather than merely averaging out performance across modalities. “Unified multimodality is the right direction, but the real difference lies in whether it can be大规模 engineered, productized, and commercialized.”
On the commercial side, C-end gross margins are too thin, and B-end models are too light.
MiniMax gross margin situation. Source: Prospectus
MiniMax Revenue Composition | Source: Prospectus
According to the prospectus, as of the first three quarters of 2025, the overall gross margin for the C-end was only 4.7%. Within the C-end revenue structure, Talkie and Xingye, focused on emotional companionship, and Hai Luo AI, focused on video creation, contributed approximately 49% and 46% of revenue, respectively.
The monetization models for these two flagship products differ. Hai Luo AI primarily operates on a subscription basis, with an average revenue per paying user (ARPPU) of $56; in contrast, users of Talkie and Xingye have weaker willingness to pay, with an ARPPU of only $5, and nearly 60% of its revenue comes from advertising.
Improvements in underlying model capabilities can directly enhance pricing power for "effect-sensitive" tools like Hai Luo AI and MiniMax App (with an ARPPU of $73), but have limited impact on ad-driven products. Hai Luo AI's scale is now approaching that of Talkie/Xingye; the nascent commercialization of agents could become a third major revenue stream and warrants close attention.
MiniMax product ARPPU (Average Revenue Per Paying User) source: prospectus
In comparison, B-end growth figures look more impressive, but the issue lies in the business model. The full-year 2025 financial report shows that enterprise services, including the open platform, generated $25.963 million in revenue, a 197.8% year-over-year increase. However, Minimax operates a “light-service” model. Founder Yan Junjie proposed at the earnings call that the value of an AI platform equals “intelligence density × token throughput.” This determines the company’s B-end strategy: scaling through token usage volume rather than relying on manpower-intensive customized projects.
According to the prospectus (as of the first three quarters of 2025), the number of paying customers on its open platform increased from approximately 100 in 2023 to about 2,500, while ARPPU decreased from approximately $12,000 in 2024 to about $6,000. This reflects a shift in its primary customer base from large enterprises to small and medium-sized businesses and developers.
For internet customers with development capabilities (such as Kingsoft Office), MiniMax only needs to provide an API interface; however, for enterprise and government clients with greater payment potential, they typically require a comprehensive suite of services including on-premises deployment, compliance assurance, and customized delivery. MiniMax’s configuration makes it better suited to serve the former type of client.
An investor following AI commented that the switching cost for large model APIs is extremely low, and selling interfaces alone makes it hard to build stickiness—eventually, one must build agents. MiniMax’s agent platform is precisely its step toward moving from “selling interfaces” to “selling outcomes.” The investor’s assessment is: “The initial experience was average, but it has improved significantly since then”—though whether it can establish a sustainable moat still needs to be proven.
MiniMax can currently be regarded as a platform company still in validation: its models have notable strengths, but it remains to be proven whether these strengths can establish stable user perception; its products have traffic, but the data flywheel and model feedback mechanisms require more time to validate; commercial growth is rapid, yet the structure and quality still offer room for optimization.
In Liu Yujia’s view, if MiniMax can demonstrate that it is a company with “model capabilities + global products + data闭环 + platform ecosystem,” its valuation potential will be substantial; however, if the market ultimately perceives it as merely a company with strong model capabilities and some overseas products, its appeal will be weaker than those companies positioned at the forefront of leading models and controlling Agent platforms, developer ecosystems, or toolchain entry points.
What awaits MiniMax after the unlock on day 03?
On July 9, MiniMax will have been listed for six months, and the first large-scale lock-up release will proceed as scheduled.
At listing, less than 3% of the total shares were available for trading on the market after excluding locked shares held by cornerstone investors, corresponding to a market capitalization of approximately HK$1.3 billion. The small float meant prices could be easily influenced by relatively small amounts of capital, pushing the stock to HK$1,330 within three months—a rise significantly driven by scarcity. On July 9, approximately 44.85% of the shares were released from lock-up, instantly expanding the free float to nearly 50%.
In contrast, Zhipu only had about 5.76% of its tokens unlocked in July, meaning its scarcity could last another six months. Starting in June, "long Zhipu, short MiniMax" became a popular paired trade.
According to the lock-up arrangements in the prospectus, these unlocked shareholders can be broadly categorized into three groups.
One category is strategic investors. Alibaba holds approximately 13%, making it the largest single external shareholder, while also serving as a cornerstone investor and additionally investing $30 million to subscribe to new shares, thereby fulfilling dual roles as both an investor and strategic partner, with relatively limited motivation for short-term liquidation. Another early shareholder, miHoYo, holds approximately 5.24% and is subject to a 12-month lock-up period, expiring in January 2027; it has also publicly stated its intention “not to reduce holdings.”
Another category consists of state-backed funds. For example, China Life, CCTV Fintech Industry Investment Fund, and Anhui Communications Control CICC—these shareholders’ investment strategies lean toward industrial positioning, and their exit timelines are typically slower. “Technically unlocked, but not in a hurry to sell,” Leo observes.
Most noteworthy are the third category: market-driven institutional investors focused purely on financial returns, such as Hillhouse, Sequoia Capital, and Future Capital. They entered early, at relatively low costs, and have achieved paper returns of several-fold to over ten-fold. For these funds, every investment has a defined exit timeline and is subject to DPI performance metrics.
Li Zeming commented: "Selling isn't really a big deal, since there's no certainty about whether we'll have the chance to achieve (higher returns)."
How much this round will drop is the market’s biggest concern right now.
According to statistics from Deren Holdings Research Institute, between 2022 and 2025, Hong Kong-listed tech stocks on average fell about 4% in the three months following lock-up expiration and about 7% in the six months. However, this is an average; Leo believes the key variable is “the distance between the entry price and the current price”—stocks that surged significantly after listing tend to experience larger declines.
Li Zeming’s assessment is relatively aggressive, stating that “it’s normal for prices to drop by half after the lock-up period ends, and there are many cases where prices have fallen 80% to 90% from their highs.”
July 9 marks the first lock-up release; three months later, MiniMax will have approximately 18.65% of its shares further unlocked. In January 2027, the 12-month lock-up periods for MiHoYo, IDG, and others will expire. The core team’s shares (founders and their close associates) are subject to a 24-month lock-up, expiring in January 2028, after which additional shares will gradually enter the market.
However, it’s important to distinguish that the unlock impact affects the stock price, not the fundamentals. MiniMax’s 2025 revenue is approximately $790 million, representing a year-over-year growth of about 160%, with a B2B gross margin of approximately 70% (as of the first three quarters of 2025); its short-term challenge lies in low C2C gross margins and an adjusted net loss of approximately $250 million for the full year, which continues to widen year-over-year; its flagship model products have just been priced 50% lower. Revenue expectations directly determine valuation.
Can the closed loop be activated? MiniMax still holds several cards: its already-launched listing on the STAR Market of the A-share market opens a new funding channel; the Agent platform represents a crucial step in shifting from selling APIs to selling outcomes; after the decline in monthly active users of Talkie, whether Hai Luo AI and the new Agent products can quickly pick up revenue will determine the sustainability of its growth. Ultimately, whether these cards can be effectively played depends on whether its technical strengths in speech synthesis, long-text processing, and other individual areas can be transformed into product moats.
Li Zeming concluded by emphasizing: If the fundamentals cannot hold up, any rebound after a decline will be weak; conversely, if the fundamentals remain strong, the period after July 9 could present an opportunity for repricing.