MiniMax, a Shanghai AI model developer founded in 2021, now serves more than 300 million users and earns most of its revenue outside mainland China. It reached global users first through consumer apps such as Talkie and Hailuo AI, and in 2026 its sales of model access to developers and businesses became its largest source of revenue.
minimax
MiniMax, a Shanghai developer of large AI models, made US$116.6 million in revenue in the first half of 2026, more than in all of 2025. About 61% of that came from outside mainland China. The company has been listed in Hong Kong since January 2026.
MiniMax first entered overseas markets with consumer apps that charge end users directly. It later opened its models to developers and enterprises. In H1 2026, revenue from developer and enterprise services surpassed revenue from its consumer apps to become its largest revenue source.
Yan Junjie co-founded MiniMax in Shanghai in December 2021. Born in 1989, he studied mathematics at Southeast University, earned a PhD in artificial intelligence at the Institute of Automation of the Chinese Academy of Sciences and completed postdoctoral research at Tsinghua University. He then spent more than six years at SenseTime, rising from intern to vice president, deputy head of the research institute and CTO of the smart city business group. He left shortly before SenseTime’s Hong Kong IPO and set up MiniMax around foundation models and AI products for consumers worldwide. Its mission is “Intelligence with Everyone”.
The company worked on text, speech and vision models at the same time rather than starting with one. Its first product, Glow, an app for chatting with AI characters, launched in October 2022 and reached several million users within months. After Glow was taken offline in China in March 2023 , its successors, Talkie overseas (June 2023) and Xingye in China (September 2023), followed and became MiniMax's first meaningful sources of revenue.
MiniMax's early backers include IDG Capital, Hillhouse, Yunqi Partners and game developer miHoYo, which invested in its US250 million Series A in mid-2023, joined by Xiaomi, Shunwei Capital and Xiaohongshu, lifting its valuation past US50 million A+ round shortly after. The largest private round came in March 2024, a US2.55 billion. Before listing, the company raised about US5.54 billion after the offer-size-adjustment and over-allotment options were both exercised in full (net proceeds about HK350 million (HK$2.72 billion).
Mainland China produced US$45.7 million of revenue in the first half of 2026, 39.2% of the total, against US$8.6 million a year earlier.
MiniMax trains its own models in four categories: text (the M series), video (Hailuo and H3), speech (Speech) and music (Music). It sells them in two ways: through its own apps, and through the Open Platform to companies and developers. In 2026 alone it has released Speech 2.8 in January, M2.5 in February, M2.7 in March, M3 in June, Music 3.0 in July and H3 at the end of July. H3 was released with open weights.
The consumer apps sold in China mirror those sold abroad. Xingye is the Chinese version of Talkie, an app in which users chat and role-play with AI characters. Hailuo AI generates video from text and images, and users worldwide had made more than 600 million videos with it by the end of 2025. H3 is now built into the app. MiniMax Agent, renamed Mavis in May 2026, handles longer multi-step tasks for individuals and teams. The apps earn from subscriptions, in-app purchases and, in Talkie and Xingye, advertising.
Most of the domestic growth has come from companies and developers. The Open Platform sells API access to MiniMax's text, video, speech and music models, and according to the prospectus Alibaba has bought API services from it since 2023. In June 2026, alongside M3, MiniMax added Token Plan, a subscription that gives developers a fixed usage allowance across its models and MiniMax Code, its coding agent. In late September the subscription was renamed M Plan and extended to include video generation. MiniMax Code's desktop application for macOS and Windows — first released in January 2026 as the MiniMax Agent desktop client and renamed MiniMax Code in late May — supports Mini Apps, small applications that users can build and run alongside the coding agent.
MiniMax's peers in China include Zhipu, which listed in Hong Kong a day before it, Moonshot AI, DeepSeek, and the model teams at Alibaba and ByteDance. MiniMax launched M3, its flagship model, on 1 June at a list price of US$0.60 per million input tokens and US$2.40 per million output tokens, with a launch discount of US$0.30 and US$1.20; the halved prices later became the standing rate for prompts of up to 512,000 tokens, making it one of the lower-cost frontier models for developers.
Overseas revenue was US$70.8 million in the first half of 2026, up from US$21.9 million a year earlier. Its share of the total fell from 71.8% to 60.8% because revenue in mainland China grew faster; in 2025 more than 70% of revenue came from abroad. In the first nine months of 2025 the company’s largest revenue sources were mainland China (27%), Singapore (24%) and the United States (20%). MiniMax says its products reach more than 300 million users in more than 230 countries and regions, and on the earnings call it said enterprise customers and developers had passed 2 million, ten times the number at the end of 2025.
MiniMax first went abroad with apps it could sell without local partners. Talkie targets young users and earns from subscriptions, in-app purchases and advertising, while Hailuo AI targets video creators, advertisers and content teams and charges by subscription tier and generation credits. The developer business came next, with M3, MiniMax Code and the Open Platform selling model access by the token. M3 has a one-million-token context window and is also available through third-party platforms such as Fireworks, and MiniMax’s own benchmark results put it ahead of GPT-5.5 and Gemini 3.1 Pro on the SWE-Bench Pro coding test.
The company is also building partnerships in individual markets. On 11 September 2026, in Shibuya, Tokyo, MiniMax and KAGAMIAI, a Japanese AI company, co-hosted the Japan IP.AI Co-Creation Conference, attended by more than 150 companies from Japan and the US. There MiniMax introduced H3 IP Edition, a version of its video model for officially licensed Japanese characters, with revenue shared with rights holders according to usage; MiniMax says more than 60 companies have since expressed interest.
MiniMax was also among the first Chinese model developers to complete registration under China’s new rules on AI technology exports. On 16 October the Ministry of Commerce and the Ministry of Science and Technology issued Announcement No. 48 of 2026, whose provisional registration requirement took effect immediately. On 23 October MiniMax said it had completed registration covering the open-source release and overseas licensing of some recent and upcoming models.
MiniMax’s revenue rose from US$30.5 million in 2024 to US$79.0 million in 2025, and reached US$116.6 million in the first half of 2026 alone; second-quarter revenue was 81.8% higher than the first quarter’s. Its Open Platform and other enterprise services brought in US$73.9 million, up 703.1% year on year, and their share of the total rose from 30.3% to 63.4%. Its AI-native products grew 100.9% to US$42.6 million. In 2025, consumer products had still made up about two-thirds of the company’s revenue.
Recurring revenue is growing faster still. On the 26 August earnings call, Yan said annual recurring revenue (ARR) had passed US$800 million in August, up from about US$150 million in February, with business customers accounting for about 80%, compared with about 30% a year earlier. Management says the figure excludes one-off licensing income; it is not audited. Token consumption on the platform in July was 20 times January’s level.
Margins improved as the company kept investing in its models. Gross margin rose from 12.1% to 17.9% year on year, which MiniMax attributes to more efficient infrastructure. Research and development spending was US$296.9 million, up 138.8%, well below the 283.1% growth in revenue, and most of it went on cloud computing for model training. Selling expenses fell 17.9% to US$27.0 million as user growth came mainly from organic channels. The net loss narrowed to US$358.0 million from US$402.2 million, and the adjusted net loss, which excludes share-based payments, fair-value changes and listing expenses, was US$293.0 million.
MiniMax held US$1.32 billion in cash at the end of June. In July a share placement and a convertible bond issue raised about HK$15.9 billion more, net, and on 29 May it began pre-listing tutoring with CITIC Securities for a second listing on Shanghai’s STAR Market. It had 568 employees at the end of June.
Demand is moving towards agents, and that favours a company selling tokens. Yan said on the earnings call that a single task given to an agent turns into many model calls, tool calls and sub-tasks, so token use is growing faster than the number of users. MiniMax is releasing models more often to keep up.
M3.1 Flash, launched on 19 October, is priced at half the rate of M3 (US0.60 per million output tokens), and MiniMax says it runs nearly three times as fast and cuts the overall cost of agent tasks by 70%. Its weights were open-sourced on 23 October; in its first 72 hours on the market it processed more than 100 trillion tokens for real users, and an early build had spent ten days topping OpenRouter's usage charts under the anonymous name "Space Bunny". In the week of 26 October to 1 November it was the eighth most-used model on OpenRouter worldwide, processing 6.23 trillion tokens, up 690% from the week before.
The model pipeline is filling out fast. According to Goldman Sachs and Citigroup notes on recent management meetings, MiniMax plans to launch M3 Pro — a model of roughly three trillion parameters built on a new attention mechanism — in early-to-mid November, followed within weeks by H3.1, an upgraded video model, and H3 Image for image generation; a new-architecture model, H4, is targeted for the first quarter of 2027, and an even larger model is already in preparation for next year. Management sees less price competition at the Pro tier, and told investors that fourth-quarter ARR had already reached the level of its US$1 billion year-end target before M3.1 Flash launched; Goldman Sachs expects ARR of about US$1.2 billion by the end of 2026, and management is targeting multiple-fold ARR growth in 2027.
Partnerships are taking a new form: instead of only selling access, MiniMax is letting its open weights become other countries' national models. In September, HUMAIN, the AI company owned by Saudi Arabia's Public Investment Fund, launched HUMAIN M3 — Saudi Arabia's first Arabic-language large model, retrained from MiniMax M3 on more than one trillion tokens of Arabic data. It averaged 89.37% across seven public Arabic benchmarks, topping five of them, and is set for sovereign deployment through the HUMAIN Node platform. M3 is also now served through Amazon Bedrock and Together AI, putting it in front of enterprise customers who never touch MiniMax's own platform.
The next stage also depends on efficiency. MiniMax's proprietary training clusters run at a 97% effective training-time rate, and it is adding domestic chips to its global compute mix. Management expects gross margin to improve in the second half of 2026 as inference costs keep falling, has said it has no near-term financing plans, and has confirmed that its planned second listing in Shanghai is proceeding as scheduled.
MiniMax shows that a Chinese model developer can earn most of its revenue abroad within five years of its founding, first through consumer apps and then by selling its models to developers and businesses worldwide.
Capital Has No Passport - In an Age of Economic Sovereignty, how to position “China” in European Capital Markets?
Oct 21, 2026 04:46 PM
Which Way Is the Wind Blowing? How to Balance Geopolitics and Heavy Capex Investment
Oct 08, 2026 05:15 PM