Zhixidongxi reported on May 30 that, according to the China Securities Regulatory Commission’s official website, MiniMax, a leading AI large model company in Shanghai, submitted its listing counseling filing report to the Shanghai CSRC on May 29, initiating its A-share listing process, with CITIC Securities serving as the counseling institution.
This also means that MiniMax will join Zhipu, which has already submitted its A-share listing guidance filing, in a race to become the first A-share company focused on large models.
MiniMax was founded in January 2022 and completed its Hong Kong IPO in January this year. Following the IPO, MiniMax’s stock price surged steadily. As of the close of trading on May 29 in Hong Kong, its stock price reached HK$840 (approximately RMB 725.24), up 409.09% from the issue price of HK$165 (approximately RMB 142.46), with a market capitalization of HK$263.454 billion (approximately RMB 227.545 billion). Starting June 8 this year, MiniMax will also be included in the Hang Seng Tech Index.
The surging stock price is supported by solid fundamentals.
On May 28, MiniMax disclosed certain business metrics. Over the past two months, MiniMax's ARR (Annualized Recurring Revenue) has grown by more than 100%, with over one million global enterprise and developer customers—a fivefold increase compared to six months ago—and a global user base of approximately 300 million.
In March this year, MiniMax released its first annual report since going public. During the earnings call, founder and CEO Yan Junjie revealed that the company’s ARR had reached $150 million as of February 2026.
In other words, combining the over 100% growth from the past two months, MiniMax’s current ARR exceeds $300 million.
In 2025, MiniMax generated $79.038 million in revenue (approximately RMB 535 million), of which $53.075 million (approximately RMB 359 million) came from AI-native products, and $25.963 million (approximately RMB 177 million) came from its open platform and other AI-based enterprise services.
Its gross profit margin increased to 25.4%, and the adjusted net loss amounted to $250 million (approximately RMB 1.69 billion), with the loss rate significantly narrowing year-over-year.
▲ MiniMax Partial Financial Data for 2025
On the product side, since the beginning of this year, MiniMax has successively launched three flagship large language models—MiniMax-M2.5, MiniMax-M2.6, and MiniMax-M2.7—and open-sourced the M2.5 and M2.7 models.
▲ Part of the MiniMax open-source models (credit: ModelScope)
With its high cost-performance ratio, MiniMax-M2.5 gained popularity among developers during the early-year "Lobster Craze" (the open-source AI agent framework OpenClaw) and was recommended in a post by Peter Steinberger, the "father of the lobster."
In mid-February this year, MiniMax briefly became the model provider with the highest market share on the AI model aggregation and routing platform OpenRouter, accounting for 18.9% of total model requests. However, MiniMax has since fallen out of the top 10 on this leaderboard.
▲ OpenRouter model provider rankings as of mid-February this year (source: OpenRouter)
In addition, in May of this year, MiniMax upgraded its Agent product and renamed it Mavis, introducing a multi-Agent parallel processing mode designed to improve the completion rate of complex, long-term tasks.
At the end of May, the MiniMax official account revealed that MiniMax-M3 is set to be released.
▲ MiniMax's official account previewed the M3 model (source: X platform @MiniMax_AI)
Skyler Miao, Head of Engineering at MiniMax, revealed additional technical details about MiniMax-M3. MiniMax-M3 employs the MiniMax Sparse Attention mechanism, achieving a 9.7x increase in inference speed during the prefilling phase when processing 1 million tokens compared to MiniMax M2. During the decoding phase, with a KV length of 1 million, speed improves by up to 15.6x, while effectively reducing attention latency.
The MiniMax Sparse Attention mechanism is built on the GQA (Grouped Query Attention) architecture. First, through the Index Branch, it computes block scores using compressed index queries (Idx Q) and index keys (Idx KV), then applies max pooling to select the top-k block indices most relevant to the current query. Next, it proceeds to the Sparse Branch, where sparse attention computations are performed only on these selected key blocks, significantly reducing computational overhead.
▲ MiniMax Sparse Attention Technical Details (Source: X platform @SkylerMiao7)
Conclusion: Leading domestic large model players rush to go public
Entering 2026, leading domestic large model companies have significantly accelerated their moves in the capital markets. In addition to MiniMax, Zhipu submitted its application for A-share listing guidance in April 2025, but ultimately completed its IPO first on the Hong Kong Stock Exchange. In February this year, Zhipu withdrew its previous A-share listing guidance application and filed a new guidance registration, adding Guotai Haitong as its new guidance advisor.
In addition, Moonshot, Jiepiao Stars, and ZeroOne AllThings have also been rumored to be planning an IPO in Hong Kong.
Faced with high computational costs and尚未完全闭环的商业化路径, leading players in large models are turning to public listings to access more diverse funding channels.
This article is from the WeChat public account "Zhi Dongxi" (ID: zhidxcom), author: Chen Junda.