Zhipu AI: China’s Generative Trailblazer Grappling with Rising Competition

Note: This post is part of our ongoing series on China’s AI unicorns. To explore the full series and learn about the other companies shaping China’s AI future, click here.

Company Background:

Zhipu AI is China’s biggest artificial intelligence (AI) start-up by the number of employees, with 800 staff, and one of its highest-valued ones, with a September 2024 valuation of approximately $2.8 billion. The company develops a range of applications from chatbots to advanced AI tools built on its General Language Model (GLM) series—large language models the company has designed to be adept at handling both Chinese and English tasks.

Zhipu was one of China’s earliest generative AI start-ups, founded in 2019 by professors Tang Jie and Li Juanzi at Tsinghua University Science Park, a sprawling innovation hub in Beijing designed to commercialize research breakthroughs and incubate high-tech ventures. While Zhipu itself stands out as a success story, its founders have had an even greater impact on China’s AI ecosystem. Juanzi, a professor in Tsinghua’s Department of Computer Science and Technology, now directs the university’s research lab—the Knowledge Engineering Group (KEG)—where Zhipu itself was first incubated and which continues to serve as a cradle of innovation for other AI startups like DeepLang and Moonshot AI (an LLM unicorn we will feature later in this series). Meanwhile, Tang, who is also a professor at Tsinghua, has personally mentored a whole generation of China’s AI leaders, including two of the founders of Moonshot AI, and serves as vice director of the Beijing Academy of Artificial Intelligence (BAAI), a leading non-profit research institution known for its development of a state-of-the-art series of language models called WuDao.

Zhipu AI, along with the contributions of its founders, played a pivotal role in jumpstarting China’s generative AI ecosystem, with its early success setting a benchmark for the industry. However, that leadership has brought new challenges. The company now faces intensifying domestic competition and the need to close the gap with global leaders like OpenAI. For now, Zhipu is holding its own. Aggressive pricing, strategic hardware partnerships, and a diversified portfolio have kept its momentum intact. But as the demands of the market evolve, sustaining this trajectory will test Zhipu’s ability to innovate at pace and maintain its position in a landscape it helped create.

Company mission:

Zhipu AI is pursuing transformative advancements in AI, aiming to lead the global race toward artificial general intelligence (AGI). Unlike incremental innovation, the company’s mission—championed by CEO Zhang Peng—is to achieve “super cognitive intelligence beyond human level.” While the aspiration aligns with global leaders like OpenAI and Google DeepMind, Zhang has emphasized Zhipu’s distinct path, built on a foundation of independent research and development.

Zhang describes the company’s journey as one of “qualitative change from quantitative change,” driven by breakthroughs in large model development. His vision of AGI involves creating systems capable of self-explanation, self-assessment, and self-supervision, underscoring Zhipu’s ambition to not only match but redefine global AGI benchmarks.

Zhipu’s July 2024 launch of technology akin to OpenAI’s Sora—a text-to-video generator hailed for its multimodal capabilities—highlights the firm’s intent to compete directly with Western leaders. A few months earlier in May 2024, Zhang acknowledged the technological gap but viewed Sora as a benchmark for innovation rather than a competitor to imitate. The pursuit of Sora-like systems, combined with Zhipu’s broader AGI goals, signals a focus on multimodal AI as a critical step in achieving AGI. Zhang is eyeing international markets as a critical milestone for the company. “Going global is a very important strategy,” he remarked in May, reflecting a vision that extends beyond domestic dominance to establishing Zhipu as a global AI leader.

What makes Zhipu’s LLMs unique?

Zhipu AI has made it clear it is trying to solve a different technical challenge than existing language models, such as Google’s BERT and OpenAI’s GPT. In a paper its researchers first submitted to arXiv in March 2021 titled “GLM: General Language Model Pretraining with Autoregressive Blank Infilling,” Zhipu AI introduced its language model GLM. While BERT, GPT, and Zhipu AI’s GLM are all based on the transformer model architecture, which was introduced in a seminal 2017 paper “Attention is All You Need,” by Google scientists, Zhipu AI contributes novel features to the LLM ecosystem by pioneering a model that is highly accurate for both Chinese and English tasks. To see what makes GLM unique, consider the following comparison:

• BERT (Bidirectional Encoder Representations from Transformers) uses a method known as masked language modeling, which essentially hides some words in the input text and trains the model to guess these hidden words using the context from both directions (before and after the hidden words). This helps BERT understand the meaning of words in context.
• GPT (Generative Pre-trained Transformer) uses an autoregressive model in which each token is predicted based on the tokens that preceded it. This is a unidirectional approach because the model is trained to predict the next word in a sentence by looking only at the words that come before it. GPT’s strength is in generating coherent and contextually appropriate text over longer passages, leveraging its training to predict the next word given all the previous words.
• GLM (General Language Model) combines features of both of these approaches by hiding parts of the text and then predicting the missing parts in sequence. This method helps the model understand the context from both directions like BERT and generate text like GPT. GLM also uses advanced techniques to handle both Chinese and English effectively, making it very versatile and powerful for a range of language tasks.

GLM is specially designed to handle Chinese with greater accuracy than other models. Unlike English script, where spaces and letters provide clear boundaries, Chinese script doesn’t have spaces, and each character represents a unique idea rather than a sound. This means LLMs have to infer word boundaries, which requires a deeper understanding of context. GLM’s training method, which analyzes text from both directions and focuses on context, is particularly suited to handle this challenge.

That said, making a powerful LLM is not only about making a model that can parse text but one that can also reason about that text. While Zhipu AI’s GLM excels in processing Chinese-language input, it seems to be behind foreign models that demonstrate advanced reasoning, decision-making, and adaptability. This is where U.S. leaders like OpenAI and Anthropic excel. Their models outperform GLM in benchmarks like SuperCLUE, which test reasoning, logic, and versatility across a range of Chinese-language tasks. The advantage likely stems from U.S. models showing stronger capabilities in higher-order logic, creative problem-solving, and general adaptability.

Indeed, in the latest November 2024 SuperCLUE rankings, OpenAI’s o1-preview holds the top position in Chinese language tasks, followed by Anthropic’s Claude 3.5 Sonnet and OpenAI’s ChatGPT-4o-latest, with Zhipu AI’s GLM-4-Plus ranking fourth. While Zhipu’s GLM is the highest-ranked Chinese-developed model, its performance lagging behind OpenAI and Anthropic highlights that it may still struggle with more complex reasoning and adaptability, areas where U.S. models have an edge.

Business model

Zhipu AI primarily operates on a Model-as-a-Service (MaaS) platform, enabling companies across various industries to fine-tune their large language models (LLMs) for specific business needs. As part of its strategy, Zhipu has adopted a selective approach to open-sourcing its GLM series, striking a balance between accessibility for developers and protecting its commercial interests. Notable open-source models include GLM-4-9B and its multimodal variant GLM-4V-9B, which offer advanced capabilities for both text and visual tasks. Additionally, GLM-4-Voice, an end-to-end speech model designed for natural human-computer interaction, has been made publicly available. However, Zhipu retains proprietary rights over certain models in the series, reflecting a calculated strategy to encourage innovation within the AI research community while safeguarding its competitive edge.

However, faced with growing competition in China’s burgeoning LLM market, Zhipu AI has sharply reduced prices for its proprietary models in a bid to secure market share. In June 2024, the company announced its second price cut within a month, slashing the cost of its flagship GLM-4 model by more than half. CEO Zhang Peng framed the move as a product of innovation rather than competitive aggression, claiming the price cuts were due to advancements in Zhipu AI’s core technologies, which he said had significantly improved efficiency and lowered costs. But Zhipu’s pricing strategy mirrors broader trends among Chinese tech giants like Alibaba, Tencent, and Baidu, which have similarly reduced the cost of their AI offerings to bolster adoption.

To better navigate the competitive LLM market, Zhipu has turned to partnerships with hardware companies such as Huawei and Qualcomm. These collaborations allow Zhipu to integrate its models with advanced chipsets and infrastructure, enabling deployments across devices like smartphones and vehicles. Through its work with Qualcomm, Zhipu integrates its AI models with chips designed to run directly on mobile devices and vehicles, making its systems faster and eliminating reliance on cloud servers. Its partnership with Huawei focuses on “AI-in-a-box” solutions—self-contained systems that let businesses run AI applications locally. This design keeps data within China’s borders and helps Zhipu’s clients comply with the country’s strict data sovereignty rules and legal requirements.

Zhipu’s offerings extend beyond models to include applications designed for both businesses and consumers. ChatGLM provides bilingual conversational AI capabilities, widely used in areas like customer support and education. It enables businesses to integrate multi-turn dialogue features into their services. AutoGLM, a personal assistant for smartphones, handles tasks such as booking appointments or placing orders through natural language commands, catering to users seeking convenience in daily activities. In addition to conversational AI, Zhipu has moved into content creation with Ying, a text-to-video tool. Ying allows users to generate short video clips from text prompts, positioning Zhipu in the growing field of multimodal AI. These applications reflect Zhipu’s effort to broaden its market reach.

Zhipu’s diversified portfolio, spanning foundational models, hardware partnerships, and application-based tools, underscores its ambition to establish itself as a leader in China’s AI market. This multifaceted approach has attracted attention not only domestically but also internationally, as foreign investors look for alternatives to Silicon Valley’s dominance in AI. Saudi Arabia’s Prosperity7, which is part of the state-owned oil group Aramco’s venture capital arm, became the first foreign investor in Zhipu AI in 2024, participating in a roughly $400 million investment round. Many consider Zhipu AI to become OpenAI’s biggest Chinese rival.

Conclusion

Zhipu AI’s rapid rise from its academic roots to a prominent force in China’s AI landscape showcases the unique potential of China’s university-driven start-up ecosystem. Zhipu capitalized on its academic foundation to attract top talent, form strategic alliances, and drive rapid innovation in the AI space. Its ambitious push toward AGI, coupled with a diversified business model, positions the company as a significant force in China’s AI ecosystem. However, as competition intensifies, Zhipu’s ability to sustain its momentum will depend on its capacity to bridge technological gaps and navigate the evolving demands of the global AI market. In this context, Zhipu stands as a key player in China’s growing AI ambitions—its success could signal a shift in the global AI landscape, challenging the dominance of Silicon Valley.