Original · Unique Research · 2025-12-26
Historical edition: This is the complete textual rendition of Unique Research’s December 2025 analysis, not an updated assessment. Prospectus figures, platform rankings, benchmark results, product claims, and litigation descriptions below are attributed to the source’s account rather than independently verified here. Its language about “making money” concerns revenue and gross/unit economics, not net profitability: the article itself reports roughly US$500 million in net losses. The US$70 million figure is annualized revenue, not a separately verified recurring-revenue measure. The source’s approximate “zero revenue two years ago” wording is retained alongside its stated 2023 revenue.
Source-consistency note: The 2024 Talkie share is described as nearly 70% of AI-native product revenue, while the later 35.1% and 32.6% figures correspond to the stated total-revenue amounts; these percentages therefore do not share an explicitly consistent denominator. The 2024 share also does not reconcile with the article’s other rounded revenue totals. The stated US$5 and US$56 per-paying-user figures have no specified measurement period here. All source figures are retained without silently resolving those gaps. “Open-source,” vendor rankings, and model comparisons retain the source’s terminology and grouping; this edition does not independently establish model-license status or benchmark comparability. The source’s link between the December 23 M2.1 release and OpenRouter rankings is preserved as its interpretation; it does not establish that their measurement windows coincide or prove causation.
MiniMax took only two years to go from zero reported revenue to an annualized US$70 million.
That pace is rare across the foundation-model industry. Rarer still, it was not sustained by enterprise contracts with a handful of major customers, but by hard cash accumulated one transaction at a time from 212 million ordinary users and 1.77 million paying users. While peers were still wrestling with "technology first or product first" and "businesses or consumers," MiniMax had already given its answer in its prospectus: it wants both, and both can make money.
On one side is Talkie, where young women pay for virtual boyfriends; on another is Hailuo AI, where creators pay for video generation; and then there is an open platform ranked sixth globally by usage on OpenRouter. These three seemingly unrelated lines, taken together, form a case study in commercializing a foundation-model company.
First, the most direct question: how is the money being made?
Two years ago, MiniMax still reported zero revenue. In 2023, that figure became US$3.46 million. By 2024, full-year revenue had jumped to just over US$30 million. The latest data is more striking: revenue reached US$53.44 million in just the first nine months of 2025, an increase of nearly 175% year over year.
In only two years, MiniMax crossed from a "zero-revenue laboratory" into an "annualized US$70 million-level commercial company." It has passed a threshold where many foundation-model companies are still struggling: it is not merely publishing papers and appearing on rankings, but bringing in money consistently.
More importantly, the quality of that revenue is improving. In 2023, overall gross margin was still -24.7%, meaning the company subsidized every transaction. But in the first nine months of 2025, overall gross margin had risen to about 23.3%, while gross margin for the open platform reached more than 70%. MiniMax is no longer simply "burning money for volume"; it has genuinely validated the unit economics of individual transactions. As the user base grows, the cost allocated to each inference continues to fall, a prerequisite for scaling a foundation-model business.
Who is paying MiniMax?
Breaking down the revenue structure, in the first nine months through the end of September 2025, AI-native products on the consumer side generated about US$38.02 million, or 71.1%, while the open platform and enterprise services on the business side generated about US$15.42 million, or 28.9%. In other words, the people paying MiniMax today are mainly ordinary users and small and midsize teams, rather than a handful of large enterprises placing large orders.
The consumer-facing business has already formed two very clear product lines, and an interesting shift is taking place.
In the public mind, MiniMax's best-known product may be the companion-chat tool Talkie, called Xingye in China. Indeed, Talkie contributed about US$19.46 million in 2024, nearly 70% of AI-native product revenue, and supported most of the business by having users buy virtual currency and view advertisements. But in the first nine months of 2025, although Talkie generated about US$18.75 million, its share fell to 35.1%.
The real change comes from Hailuo AI. This video-generation tool generated only about US$2.35 million in 2024 and was still the "little brother." But in the first nine months of 2025, Hailuo AI's revenue surged to about US$17.46 million, representing 32.6% and nearly drawing level with Talkie. This curve closely matches the app data for November: Hailuo AI recorded 540,000 downloads that month, up 19.19% month over month, while the web version of MiniMax's official site received 1.97 million visits, soaring 143.82% month over month.
Why was Hailuo able to take over growth so quickly? Average revenue per paying user makes the answer clear. Talkie's paying users contribute about US$5 per person, while Hailuo AI's contribute US$56, more than ten times as much. Users may chat with virtual people for "companionship," but they are more willing to pay for "productivity." Xingye helped MiniMax earn its first pot of gold; Hailuo is becoming its second growth curve.
Judging by revenue and customer structures, MiniMax and Zhipu are following two diametrically opposed paths. More than seventy percent of MiniMax's revenue comes from the consumer side, its customers are highly dispersed, overseas revenue exceeds 70%, and overall gross margin is only 23%: a classic "small margin, high volume" strategy. Zhipu, by contrast, derives 85% of revenue from localized deployment, or private deployment, and serves 12,000 institutional customers, but its five largest customers contribute nearly half of revenue, its largest customer accounts for more than 10%, and gross margin remains around 50%: an asset-heavy model of "higher margins through focused execution." Put simply, MiniMax earns nimble subscription and advertising revenue, while Zhipu earns laborious customization revenue. The former resembles a consumer-goods business; the latter resembles project-based contracting.
Developers Vote with Their Feet
If the consumer side packages models directly into products and sells them to ordinary users, the open platform sells capabilities to developers and enterprises. This business receives less attention than Talkie and Hailuo, but its figures are solid: about US$2.70 million in 2023; rising to about US$8.72 million in 2024; and about US$15.42 million in the first nine months of 2025. Gross margin has remained around 60-75% for years, typical of an API business: the cheaper the computing power and the greater the scale, the better the gross margin.
According to the State of AI report jointly published by OpenRouter and a16z, MiniMax has consumed about 1.26 trillion tokens on OpenRouter by token-call volume, ranking sixth globally among open-source model vendors. Ahead are DeepSeek, Qwen, Meta LLaMA, Mistral, and OpenAI; Moonshot and Google come after it.
In other words, among highly pragmatic developers worldwide, MiniMax is not a fringe player; it has entered the leading open-source/open-model group. Developers are intensely practical and vote with their feet, going wherever the price-performance ratio and usability are best. Consumer-facing products prove that "this model can attract ordinary users," while the business-facing open platform proves that "this model can also fit into other people's product stacks." Only by combining the two does the model truly become infrastructure.
User Scale: Not the Story of a Niche Community
User data shows that MiniMax has also achieved substantial scale. As of September 30, 2025, its products had cumulatively reached about 212 million users; AI-native products averaged about 27.62 million monthly active users; and nearly 100 million new users registered in the first nine months of 2025 alone.
●Breaking down the core products: Talkie/ Xingye had about 147 million users and roughly 20 million monthly active users;
●Hailuo AI had about 42.35 million users and roughly 5.61 million monthly active users;
●MiniMax's own Agent/ office products had also exceeded 19 million users.
Paying users are what truly determine commercial value. In the first nine months of 2025, consumer-facing products had about 1.77 million cumulative paying users, including approximately 1.39 million for Xingye; 310,000 for Hailuo AI; and 70,000 across other products. The open platform's paying business-facing customers, meanwhile, grew from about 100 in 2023 to roughly 2,500.
Combined with ARPU data, a clear structure emerges: Talkie is a low-ticket-price, massive-scale "companionship and emotion" business; Hailuo is a high-ticket-price, fast-expanding "productivity and monetization" business; and the open platform is a high-margin infrastructure business compounded by developers. With these three pieces together, MiniMax is no longer a company "betting on a single super app," but a layered commercial portfolio.
The Technology Foundation Determines the Commercial Ceiling
Many people ask: why has MiniMax taken off simultaneously in commercialization and developer reputation? The answer is not merely that "the model is powerful," but lies in its technical path over these two years.
MiniMax M2.1, released on December 23, is a sparse model with 10 billion active parameters, tailored to real-world coding and agent scenarios. It scored 72.5% on SWE-multilingual and 88.6 on its own open-source VIBE-bench, placing in the first tier on Web, Android, and other subcategories. In multilingual programming, it achieved "open source overtaking closed source" on a number of subcategories compared with closed models such as Gemini 3 Pro and Claude 4.5 Sonnet.
More importantly, M2.1 is already available via API on the MiniMax open platform; its weights are fully open source for local inference, and it is embedded in products such as MiniMax Agent, supporting a complete workflow spanning web development, app development, and long-chain office automation. This directly explains why MiniMax appears in the OpenRouter report's top six and why the designation M2/M2.1 keeps appearing in various IDE plug-ins and agent tools.
On the other side, Hailuo AI has turned MiniMax's multimodal and video-generation capabilities into a product line highly attractive to creators. The Hailuo-2.3 series now ranks steadily among the global top 10-20 and the top 5 Chinese models on Text-to-Video, Image-to-Video, and other leaderboards. The new Hailuo-2.3 works together with M2.1 to strengthen character movement, expressions, physical details, and text-to-video consistency; combined with a fast version, it gives creators two settings, "quality" and "efficiency."
Behind this is a product logic: at the top are consumer-facing content tools such as Hailuo, in the middle is the business-facing open platform or API, and at the bottom are open-source models represented by M2.1. Technology forms the foundation, products form the shells, and commercialization flows among those shells. That is MiniMax's current basic structure.
Continued Heavy Investment in R&D
Of course, there is a practical problem behind the impressive data: MiniMax continues to record large losses. Its prospectus states that over the three years from 2022-2024, R&D expenditure surged from about US$10.60 million to nearly US$189 million; another approximately US$180 million was invested during the first nine months of 2025; and the net loss for the first nine months of 2025 was around US$500 million, mainly due to R&D and changes in the fair value of related financial liabilities.
In other words, this is a company that has proven it "can make money" but continues to "bet heavily on R&D" to build stronger models. MiniMax also said in its prospectus that the core keyword for the future is "scalability": using technologies such as MoE, or mixture of experts, architecture and linear attention to drive down inference costs, while using model families tailored to specific scenarios, such as M2.1 and Hailuo, to raise model capabilities.
Can it find a genuinely sustainable balance between the cost curve and the capability curve? This is not only MiniMax's question, but one the entire Chinese foundation-model sector and even the global agent ecosystem must answer together.
The Copyright Risk Hanging Overhead
As Hailuo AI scales rapidly, MiniMax has also encountered its first compliance test. The prospectus discloses that several US film companies, including Disney and Universal Pictures, have filed copyright lawsuits over Hailuo-generated content, alleging that its output videos infringe rights in their works. MiniMax assesses the "risk as controllable" and has estimated a potential upper limit on damages.
It is like a sword hanging overhead. In the short term, it will indeed affect the pace of Hailuo's business expansion in some markets. In the long term, however, as long as generative video continues to enter mainstream production relationships, such lawsuits are almost inevitable: whoever runs in front explores the legal boundaries first. For MiniMax, what matters more is how to preserve the product's "creative freedom" while proactively building a "copyright protection net" at the technology and product levels before compliance boundaries become clear. This will directly affect how far the Hailuo business can go.
A Commercialization Case Study for a Chinese Foundation Model
Taken together, all the numbers, rankings, and stories make MiniMax a highly representative case study of a Chinese foundation-model company. It first used consumer-facing super apps to generate cash flow: Xingye built scale through "emotional companionship," while Hailuo used a "productivity tool" to drive average spending. This allowed the model to complete market validation on the consumer side before long-cycle contracts from large enterprises arrived.
It then used the open platform to turn the model into "infrastructure inside other people's applications." The 1.26 trillion tokens and sixth-place global ranking in the OpenRouter report show that it has entered developers' mainstream consideration set. For entrepreneurs, this is a typical path of "building the brand through the consumer side first, then compounding returns through the business side."
The technology, product, and commercial loops all move fast enough. From M2 to M2.1 and then Hailuo-2.3, the company is catching up with, and in some dimensions surpassing, overseas closed models while delivering strong user and growth data on App/Web rankings.
The OpenRouter report mentions a "Cinderella glass-slipper effect": when a model is the first to precisely address a particular pain-point scenario and find the "glass slipper that fits best," users will remain in those shoes for a long time.
●In companion chat, Talkie put that glass slipper on its own foot;
●In video creation and multilingual coding, Hailuo and M2.1 are trying to put on another pair.
For people following the international expansion of Chinese AI, agent commercialization, and model ecosystems, the greatest value of MiniMax's prospectus is not "how much it sold," but its demonstration to the market of a path that can be replicated and also deserves questioning and refinement: how, amid computing pressure, cutthroat competition, and tightening regulation in 2025, a foundation model can be turned from a "technology sample" into a truly sustainable business.
What remains to be seen is whether, after M2.1 becomes open source, Hailuo continues climbing, and global copyright rules gradually become clearer, MiniMax can travel a little farther along this path than the numbers in its current prospectus suggest.