Investment Research Team
Zhipu AI, a generative AI startup in China originating from Tsinghua University, unveiled several new products at its Open Day event in Beijing. These include the GLM-4 series of open-source large models, new API models, the MaaS 2.0 platform, and various intelligent agents.
Following a price cut in May, Zhipu AI has announced further price reduction yesterday for its AI services amid a wave of price cuts among similar AI service providers in China.
Founded in 2019 and headquartered in Beijing, Zhipu AI recently achieved a US$3 billion valuation after receiving a US$400 million investment from Prosperity 7, a fund managed by Middle Eastern oil giant Aramco’s venture capital arm. With over 800 employees, the company serves industries such as media and finance.
Will significant investments and competitive pricing position Zhipu AI as China’s answer to OpenAI among numerous similar startups? Let’s delve deeper into Zhipu AI’s journey and potential.
Zhipu AI is an AI company specializing in large language models. CEO Zhang Peng graduated from the Department of Computer Science and Technology at Tsinghua University, with his doctoral research focused on knowledge graphs. Chairman Liu Debing previously served as Deputy Director of the Technology Big Data Research Center at Tsinghua Data Science Research Institute. President Wang Shaolan holds a Ph.D. from Tsinghua University.
It’s products include GLM-4 series of large models, an open platform for AI services, and the Zhipu Qingyan App, which hosts over 300,000 active intelligent agents for productivity tasks.
The company has raised several rounds of financing from notable investors such as Beijing State-owned Capital Operation and Management, Lightspeed China Partners, Alibaba Cloud Computing, Ant Group, and GL Ventures. After receiving US$400 million from Prosperity 7 last week, it’s valuation has reportedly reached US$3 billion.
Currently, Zhipu AI’s open platform serves 300,000 enterprise customers, with a daily call volume reaching 40 billion tokens. Over the past six months, API daily consumption has increased more than 50 times, according to information shared during its Open Day.
Zhipu AI’s consumer-facing product, the Zhipu Qingyan App, now hosts over 300,000 active intelligent agents in its intelligent agent center, including productivity tools such as mind mapping, document assistants, and scheduling.
The new version of the MaaS large model open platform offers a more segmented pricing structure, providing multiple packages for enterprise customers to choose from. Without needing any code, customers can complete fine-tuning in three steps and deploy private models using their own data.
After the latest price cut, Zhipu AI’s lightweight version of GLM-4-Flash is costing as low as 6 cents per million tokens.
Under the enterprise V3 pricing, generating an image with the CogView-3 text-to-image model costs only 6 cents.
Zhipu AI also open-sourced a 9-billion-parameter GLM-4-9B model, which boasts a 40% improvement in overall capabilities compared to ChatGLM3-6B and outperforms Llama-3-8B-Instruct, according to the company’s own information.
In comparison to the Llama-3-8B model, which has a larger training dataset, the GLM-4-49B shows a slight advantage in English and a 50% improvement in Chinese academic capabilities, the company claims.
The context window has also been expanded from 128K to 1M tokens. This means the model can process 2 million characters of input simultaneously.
The GLM-4-9B model also possesses the ability to call external tools, such as code execution, web browsing, drawing, file operations, database queries, API calls, etc., to assist in answering questions or completing tasks.
In terms of model size, the GLM-4-9B is a small-sized model suitable for various application scenarios. Zhipu AI emphasized that they did not aim to increase the number of parameters in their open-source models.
Zhipu AI’s MaaS large-scale model open platform has released version 2.0. In contrast to the January Dev Day, which highlighted the GLM-4’s capabilities comparable to GPT-4, the new MaaS platform focuses on commercial applications with various package options:
GLM-4 costs about 100 yuan/1M tokens
GLM-4-Air offers high performance and is cost-effective
GLM-4-Flash is the most affordable, costing 0.1 yuan/1M tokens
CogView-3 can generates images for 0.1 yuan each.
There are certainly many generative AI products from various companies in China. The competitiveness of Zhipu AI’s offerings remains to be seen.
Related News:
