2026 IPO bonanza! Over 90% of new stocks rose on their debut

钛媒体APP

joined discussion ·

Zhipu and MiniMax: Two Different 'Approaches' to Large Models

By Bo Hu Finance, Author: Kaikai

On January 8, Zhipu, the world's first publicly traded large model company, officially listed on the Hong Kong Stock Exchange, followed shortly by another large model unicorn, MiniMax. The flurry of IPOs within two days has brought China’s AI large model commercial narrative to the forefront.

As the 'twin titans' of domestic large models, the listing progress of Zhipu and MiniMax has been closely watched. Both are top players among the 'Six小龙 (Little Dragons)' of China’s AI sector, with pre-IPO valuations of around HKD 50 billion each, both setting new records for subscription热度 (heat) in recent AI IPOs.

As of the date of writing, Zhipu and MiniMax have market capitalizations of approximately HKD 80 billion and HKD 110 billion, respectively, further widening the valuation gap from before their listings. Among them, Zhipu has adopted a more steady approach, while MiniMax has garnered greater热度 (attention).

While the enthusiastic response from the capital markets is eye-catching, the biggest悬念 (suspense) going forward will be whether the distinct strategic genes and commercialization paths of these two large model unicorns can offer a verifiable and replicable development paradigm for China's large model industry.

As two AI companies founded before the ChatGPT boom, Zhipu AI CEO Zhang Peng and MiniMax founder Yan Junjie are both staunch believers in AGI (Artificial General Intelligence).

But the darkness before dawn is always prolonged, and maintaining faith during times when no one understands requires enduring more 'lonely moments.'

Zhang Peng was rooted in Tsinghua University’s Computer Science Department Knowledge Engineering Lab (KEG), which launched the scientific information analysis engine AMiner in 2016. In 2019, Zhang Peng led his team to break away independently and officially founded Zhipu AI.

Coincidentally, Yan Junjie also conducted postdoctoral research at Tsinghua University’s Computer Science Department. During his Ph.D., he interned at Baidu's AI Research Institute, later joining SenseTime, rising from an intern to Group Vice President, where he led the construction of deep learning toolchains and general intelligent technology systems.

It can be seen that before becoming founders of two unicorn enterprises, Zhang Peng and Yan Junjie had already accumulated significant artificial intelligence technology expertise in their respective fields, allowing them to foresee the AI era's trends earlier.

As the clock struck 2021, artificial intelligence was still a quiet track in the corner, while the hottest entrepreneurial track was the metaverse, which even Facebook had to rename its corporate identity for.

However, in the same year, Zhipu had already secured RMB 100 million in Series A funding, officially transitioning from a 'research laboratory' to the 'commercialization track for large models,' and by August 2022, developed one of China’s earliest large models with hundreds of billions of parameters.

In 2022, driven by the mission of 'letting ordinary people feel the changes brought by AI,' Yan Junjie left his position as Vice President of SenseTime and founded MiniMax, heading towards the goal of AGI (Artificial General Intelligence).

Yan Junjie believed that AI designed for specific tasks had already hit its commercial ceiling, whereas general artificial intelligence does not require customization and can form standardized products to serve massive numbers of users—this is the ultimate direction for creating scalable value.

Zhang Peng's entrepreneurial philosophy carries a stronger sense of 'idealism.' He frankly admitted that the original intention of founding Zhipu was not purely profit-driven but aimed at advancing work within the industry, aligning more closely with the current needs of artificial intelligence development.

The rest of the story is well-known. In early 2023, ChatGPT-4 was officially released, and large AI models quickly became a technological wave that even ordinary people could perceive.

Following this, major domestic internet companies rolled out their self-developed large models, and startups like MoonShot and 01.AI also sprouted up like mushrooms after rain. Zhipu and MiniMax began to feel more urgent commercial pressures.

From this point on, Zhipu and MiniMax gradually went down diverging paths.

As a company born out of Tsinghua University's lab, Zhipu, with its 'solid roots,' has no shortage of state-owned investment backing, such as Hangzhou Urban Investment, Shangcheng Capital, and Zhuhai Huafa.

Under the backdrop of deep integration with state-owned enterprises, Zhipu has also gained more government and enterprise clients as well as academic clients, such as developing 'AI + academic search products' for universities and research institutions, and the 'Zhipu + Zhuhai Huafa Space' project developed in collaboration with Zhuhai Huafa.

If Zhilv is considered an unequivocal 'academic school,' then MiniMax resembles more of a 'wild card' that thrives in diverse environments.

From the perspective of its financing background, MiniMax is supported by major tech companies and international investors, such as Alibaba, Tencent, Xiaohongshu, as well as the Abu Dhabi Investment Authority and Mirae Asset of Korea, which have opened up vast overseas markets for MiniMax.

Yan Junjie places greater emphasis on the generalization capabilities of large models. He believes that a strong technology company should not merely be a seller of technology but a product-driven technology firm. This philosophy has also led MiniMax to focus more on the C-end market.

With different backgrounds and resources, these two AI unicorns have adopted vastly different technical paths and commercialization strategies.

Zhipu chose to independently develop the GLM (General Language Model) path, which differs from OpenAI's GPT architecture. Zhang Peng once stated that domestic large models must achieve their own innovation and should not simply replicate the technical paths of world-leading levels.

The GLM architecture can utilize both forward and backward information when processing language tasks, giving it advantages in long-text comprehension, logical reasoning, and lower hallucination rates. Zhang Peng mentioned that theoretically, GLM's training efficiency would be higher than GPT's. Additionally, Zhilv’s commitment to a fully self-controlled technical route has helped it gain unique trust in the government and enterprise market.

According to the prospectus, Zhipu has become one of China's largest independent large model vendors by revenue scale. From 2022 to the first half of 2025, it achieved revenues of RMB 57.4 million, RMB 124.5 million, RMB 312.4 million, and RMB 190 million respectively, with a compound annual growth rate reaching 130%.

（图源：智谱招股书）

The company primarily provides localized deployment and cloud API services through its MaaS (Model as a Service) platform. In 2024, localized deployment revenue accounted for 84.5% of Zhilv’s total revenue, mainly targeting institutional clients such as government, finance, and energy sectors that have high data security requirements, offering them customized model solutions.

（图源：智谱招股书）

By focusing on a B-end oriented business model, Zhilv has secured more stable and higher gross margin revenue streams, maintaining a gross margin consistently above 50% over the long term.

However, the flip side of this business model, which focuses on localized deployment, also brings challenges such as long delivery cycles, high customization demands, high customer concentration, and difficulties in scaling, making it hard to support long-term sustainable growth.

To break through this bottleneck, Zhipu is also accelerating its transition to the cloud. The proportion of its cloud deployment business in total revenue has increased from 4.5% in 2022 to 15.5% in 2024. Zhang Peng stated that he hopes to increase the revenue share of API business to 50%.

Unlike Zhipu’s commercialization route of being the 'computing power infrastructure' for the AI era, MiniMax has embarked on a path of going global with C-end products.

Yan Junjie's viewpoint is very clear: he believes that no matter how good large model technology is, there must be sufficiently strong product capabilities to support it. This “model as a product” concept bears a resemblance to ByteDance's 'APP factory' strategy.

But for MiniMax, which started relatively late, catching up is the key challenge. To address this, Yan Junjie is betting on the MoE (Mixture of Experts) hybrid model, which allows the model to become more complex and powerful without significantly increasing computational resources.

MiniMax revealed that its next-generation model, based on MOE + Linear Attention, can improve efficiency by 2-3 times when processing 100,000 tokens compared to the same generation of models like GPT-4o, with greater improvements seen as the length increases.

This lightweight design gives MiniMax a cost advantage in the high computing power-consuming C-end market. Currently, MiniMax's representative products include 'Conch AI' with text-to-video functionality and 'Starry/Talkie,' an AI emotional companion app.

As of the first three quarters of 2025, over 70% of MiniMax’s revenue comes from overseas, with C-end AI applications contributing approximately 71% of income, primarily adopting a monetization model of 'advertising + subscription + in-app purchases.'

The popularity of its C-end applications has also helped MiniMax accumulate a substantial user base. Its AI-native product matrix boasts an average of 27.6 million monthly active users, with cumulative users exceeding 212 million, of which just the Talkie/Xingye product alone has gathered 147 million users.

（图源：MiniMax招股书）

Over the past three years, MiniMax's revenue has also shown explosive growth. In the first three quarters of 2023 to 2025, the figures were $3.5 million (approximately RMB 24.644 million), $30.5 million (approximately RMB 220 million), and $53.4 million (approximately RMB 380 million), respectively.

（图源：MiniMax招股书）

However, although MiniMax’s revenue growth far exceeds that of Zhipu, the vulnerability of the C-end model is also obvious. MiniMax’s gross profit margin has long been lower than Zhipu’s, with the core C-end business gross profit margin reaching only 4.7% in the first three quarters of 2025.

For MiniMax, how to break through the ceiling of user growth, continuously enhance users' willingness to pay, and address regulatory risks in overseas markets will become key to verifying whether its “product-driven” model can succeed.

After six years of evolution, Zhipu and MiniMax have evolved into two distinct entities—one heading left, the other right, seemingly becoming parallel lines.

However, both companies launched their assaults on the capital market almost simultaneously and were inevitably placed under the same spotlight, facing a common reality: large-scale model enterprises are still incurring losses.

From 2022 to 2024, Zhipu’s net losses were RMB 144 million, RMB 788 million, and RMB 2.958 billion, with a loss of RMB 2.358 billion in the first half of 2025; during the same period, MiniMax’s net losses were $73.7 million, $269 million, and $465 million, with a loss of $512 million in the first three quarters of 2025. Both companies’ losses continue to widen, and they still lack self-sustaining profitability.

For Zhipu and MiniMax, the main areas of expenditure remain the research and development of large-scale models and computing power. In 2024, Zhipu’s R&D investment reached RMB 2.195 billion, seven times its revenue that year; MiniMax’s R&D spending was also about six times its revenue. For both companies, 70%-80% of their R&D investment went toward computing power.

In addition, MiniMax invested heavily in marketing and user acquisition in global markets during its early stages. In 2024, its sales and marketing expenses accounted for 285% of total revenue. However, MiniMax’s sales and marketing expenditures declined last year. Yan Junjie publicly stated that he is unwilling to invest significant amounts of money in unproductive traffic acquisition and prefers to attract users through the inherent capabilities of the product itself.

Thus, although Zhipu and MiniMax both exhibit the common characteristics of 'high growth, high investment, not yet profitable,' the stories they tell are different, and the capital market also shows varying attitudes.

Based on the subscription situation before the IPO, MiniMax's popularity was significantly higher than Zhipu's. On the first day of listing, Zhipu’s price once fell below its issue price during trading, closing 13.17% higher than its offering price of HK$116.2 per share; MiniMax closed over 109% higher than its offering price of HK$165 per share.

An investor focused on AI mentioned that in the Chinese market, investors generally prioritize C-end companies over B-end ones. MiniMax has told a story centered around C-end products, which have a higher ceiling for user scale and business model; Zhipu’s growth is predictable but often lacks the explosive speed of C-end models, making the former more favored by the capital markets.

However, at this stage, both companies are in the early phases of their IPOs, actively addressing and optimizing their respective operational weaknesses. It is difficult to determine at present which path will ultimately prove more correct.

What is certain, though, is that the IPOs of Zhipu and MiniMax provide clearer reference examples for domestic large model enterprises in China, shifting the focus of the industry from pure 'technological narrative' to 'business narrative.'

Faced with the technological lead of foreign large models and the saturated offensive from domestic tech giants, startups in the large model space must find their own differentiated paths in their business models if they want to secure a place at the table.

While technology-driven innovation is crucial, even more important is creating a product that customers are willing to pay for, rather than merely a good model. China's AI unicorns must navigate beyond the shallows of technology and dive deep into the ocean of capital to hone their true business capabilities.

Therefore, going public is not the end for large model enterprises, but rather an opportunity and challenge to showcase themselves more fully.

Under the scrutiny of global investors, whether it is Zhipu or MiniMax, or subsequent large model enterprises that can step onto the capital stage, they all must answer one key question: after burning money for the future, who can truly survive into the future.

Large model enterprises have limited 'ammunition,' and as the market continues to expand, scale advantages may turn into cost pressures. At that point, whoever can build a lighter model, reduce computing power costs further; accomplish the same tasks with less computing power; or be the first to run through a viable business model, spreading out computing costs, will become the new competitive focal points.

The path to AGI is bound to be long, but the current global large model industry is no longer willing to just listen to 'stories'.

336K Views