Zhipu AI
Last reviewed
Apr 28, 2026
Sources
14 citations
Review status
Source-backed
Revision
v7 · 4,168 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Apr 28, 2026
Sources
14 citations
Review status
Source-backed
Revision
v7 · 4,168 words
Add missing citations, update stale details, or suggest a clearer explanation.
Zhipu AI (智谱AI), now branded internationally as Z.ai, is a Chinese artificial intelligence company headquartered in Beijing. Founded on June 11, 2019, the company originated as a spin-off from the Knowledge Engineering Group (KEG) at Tsinghua University. Zhipu develops the GLM (General Language Model) family of large language models, multimodal models, and AI agent systems. In January 2026, Zhipu became the first Chinese AI foundation model company to go public, listing on the Hong Kong Stock Exchange under the ticker 02513 and the legal entity name Knowledge Atlas Technology.
Often described as "China's OpenAI," Zhipu is recognized as one of China's "Six Little Tigers" (六小虎), a group of prominent AI startups that emerged during the generative AI wave. By the time of its IPO, the company had raised approximately $1.5 billion in total funding from investors including Alibaba, Tencent, Meituan, Xiaomi, HongShan (formerly Sequoia China), and Saudi Arabia's Prosperity7 Ventures.
Zhipu AI was established on June 11, 2019, by Tang Jie and Li Juanzi, both professors in the Department of Computer Science and Technology at Tsinghua University. The company grew out of the university's Knowledge Engineering Group (KEG), a research lab focused on knowledge graphs, natural language processing, and social network analysis. Tang Jie, a Fellow of the ACM, IEEE, and AAAI, had led the KEG for years and mentored a generation of Chinese AI researchers, including several founders of rival AI startups. Tang also serves as vice director of the Beijing Academy of Artificial Intelligence (BAAI), an institution that played a role in early Chinese large language model research.
The company was initially located in the Tsinghua University Science Park in Beijing's Zhongguancun district, a technology hub often compared to Silicon Valley. In its earliest days, Zhipu struggled to secure funding as one of many academic spin-offs without a proven business model. The administrative commission of Zhongguancun Science Park provided the founding team with rent-free office space for three months, giving them room to develop their initial products. The company's first commercial focus was building knowledge graphs to support enterprise research and innovation workflows.
In 2020, even before the wider large language model boom triggered by the success of GPT-3, Tang Jie and his team began investing in large-model AI research alongside their existing knowledge graph business. At a time when most Chinese AI companies were focused on computer vision or speech recognition, this early bet on foundation model technology would prove to be a defining strategic decision that set Zhipu apart from its peers.
During this period, the team also published CogView, a 4-billion-parameter Transformer-based text-to-image model, presented at NeurIPS 2021. CogView used a VQ-VAE tokenizer and achieved state-of-the-art FID scores on the blurred MS COCO dataset, outperforming prior GAN-based approaches. It was among the first Chinese models to compete with OpenAI's DALL-E in image generation from text descriptions.
In August 2022, Zhipu and Tsinghua's KEG released GLM-130B, a bilingual (English and Chinese) pre-trained language model with 130 billion parameters. This made Zhipu the developer of China's first self-developed pre-trained language model of this scale. The model was published as a paper that was accepted at ICLR 2023 ("GLM-130B: An Open Bilingual Pre-trained Model"). GLM-130B used the General Language Model framework, which combines autoregressive blank infilling with bidirectional attention, a unique approach distinct from both the GPT-style autoregressive paradigm and the BERT-style masked language modeling approach.
GLM-130B was trained on approximately 400 billion tokens (roughly 200 billion each in English and Chinese) and featured 70 Transformer layers with a hidden dimension of 12,288 and a bilingual tokenizer containing 150,000 tokens. The pre-training objective consisted of 95% self-supervised training using autoregressive blank infilling over the public Pile corpus for English and several Chinese corpora, plus 5% multi-task instruction pre-training across 70 different datasets.
On benchmarks, GLM-130B outperformed GPT-3 175B on LAMBADA (+5.0%), OPT-175B (+6.5%), and BLOOM-176B (+13.0%). It performed comparably to GPT-3 175B on MMLU (+0.9%). For Chinese-language tasks, it significantly exceeded ERNIE TITAN 3.0 260B on zero-shot CLUE datasets (+24.26%) and FewCLUE datasets (+12.75%). The model could run inference on a single server with 8 A100 (40GB) or 8 V100 (32GB) GPUs, making it relatively accessible for a model of its size.
Also in 2022, CogView2 was released, a hierarchical Transformer-based text-to-image model that achieved a 10x speedup over the original CogView while producing competitive image quality against DALL-E 2.
These releases established Zhipu as a serious competitor in the global large language model landscape and attracted growing investor interest.
On March 14, 2023, Zhipu open-sourced ChatGLM-6B, a 6.2-billion-parameter conversational model derived from the GLM architecture. This lightweight model was optimized for Chinese question-answering and dialogue. It was trained on approximately one trillion tokens of Chinese and English corpus, supplemented by supervised fine-tuning, feedback bootstrap, and reinforcement learning from human feedback (RLHF). The model could run on consumer-grade GPUs using INT4 quantization with as little as 6 GB of GPU memory. The release attracted massive attention in the Chinese developer community, accumulating over 10 million downloads on Hugging Face within its first year and becoming one of the most popular open-source language models globally during 2023.
ChatGLM-6B was quickly followed by ChatGLM2-6B in June 2023, an updated version with stronger benchmark performance, extended context length (from 2K to 32K tokens), and more efficient inference (a 42% speedup over the original).
On October 22, 2023, Zhipu released ChatGLM3-6B, a third-generation model that natively supported function calling, code interpretation, and agent-based workflows in complex scenarios. ChatGLM3 was released in multiple variants: ChatGLM3-6B (standard), ChatGLM3-6B-Base, ChatGLM3-6B-32K, and ChatGLM3-6B-128K. With the ChatGLM3 launch, Zhipu's consumer chatbot product, Zhipu Qingyan (智谱清言), became the first large model product in China with built-in code interaction capability. Also at the October 2023 China National Computer Conference, Zhipu open-sourced CogVLM-17B, a multimodal vision-language model.
Also in October 2023, Zhipu secured a major funding round of approximately $342 million with participation from Alibaba, Tencent, and other major Chinese tech firms, valuing the company at around $2.8 billion. This round represented one of the largest single investments in a Chinese AI startup at the time.
In January 2024, Zhipu released GLM-4, its fourth-generation base model. GLM-4 represented a 60% performance improvement over GLM-3 and achieved scores comparable to GPT-4 on general benchmarks including MMLU, GSM8K, MATH, BBH, GPQA, and HumanEval. The model supported a 128K token context window and achieved 100% accurate recall in needle-in-a-haystack testing at that context length. It also scored up to 90% on the IFEval instruction-following benchmark. GLM-4 was pre-trained on over ten trillion tokens across Chinese, English, and 24 additional languages.
The open-source GLM-4-9B-Chat variant offered features such as web browsing, code execution, custom tool calling (function call), and long-context reasoning for the broader developer community.
In May 2024, Zhipu raised approximately $400 million in a round that included Saudi Arabia's Prosperity7 Ventures, marking the company's first foreign investor participation. The round valued Zhipu at roughly $3 billion.
In June 2024, Zhipu slashed GLM-4 API prices by over 50%, joining a broader price war among Chinese AI providers that saw providers such as Alibaba (Qwen), Baidu (ERNIE), and ByteDance (Doubao) aggressively cutting prices to attract developers.
Later in 2024, Zhipu released GLM-4-Plus, a next-generation proprietary model with performance comparable to GPT-4o. In Chinese-language benchmarks, GLM-4-Plus ranked fourth in the SuperCLUE evaluation as of November 2024. The company also launched GLM-4V, a multimodal vision-language model comparable to GPT-4V, and GLM-4-Voice, a model with real-time voice interaction capabilities that brought a "video call" feature to the Zhipu Qingyan app.
In November 2024, Zhipu introduced AutoGLM, a smartphone AI agent capable of executing multi-step tasks such as ordering food delivery, booking flights, and navigating apps by interpreting on-screen content and simulating human-like taps, swipes, and text input. AutoGLM could handle operation sequences exceeding 50 steps across multiple applications and supported browser-based website automation.
In December 2024, the company completed a Series D funding round of approximately $420 million, its largest single round to date.
In April 2025, Zhipu pre-filed for an initial public offering on the Hong Kong Stock Exchange, becoming the first of China's "Six Little Tigers" to pursue a public listing. Shortly after, the company open-sourced the GLM-4-0414 model series under the MIT license, including a 32-billion-parameter base model (GLM-4-32B-0414) pre-trained on 15 terabytes of high-quality data, a reasoning model (GLM-Z1-32B-0414) that used cold-start and extended reinforcement learning strategies, and a rumination model (GLM-Z1-Rumination-32B-0414) designed for multi-step deliberation on complex problems. The reasoning model, with only 32 billion parameters, achieved performance on certain tasks comparable to DeepSeek-R1, which has 671 billion parameters.
In July 2025, the company released GLM-4.5 and GLM-4.5 Air and formally rebranded its international identity as Z.ai while retaining the Zhipu AI (智谱AI) name in the Chinese market. Usage of the subscription service grew tenfold in the two months following the GLM-4.5 launch.
Throughout 2025, Zhipu also received several rounds of government-backed investment. In March 2025, the company raised approximately $257 million from funds backed by the cities of Chengdu, Zhuhai, and Hangzhou. In July 2025, Shanghai state funds contributed an additional $140 million.
In December 2025, Zhipu open-sourced the GLM-4.6V series, a pair of vision-language models with native function-calling capabilities. GLM-4.6V (106B parameters) targeted cloud deployments, while GLM-4.6V-Flash (9B parameters) was designed for local, low-latency use. Both supported a 128K-token training context window, enough to process approximately 150 pages of dense documents or one hour of video in a single pass.
Also in December 2025, Z.ai open-sourced GLM-4.7, a large language model built for real development workflows, including coding tasks. The company also released CogKit, a fine-tuning and inference framework for the CogView4 and CogVideoX model families.
On January 8, 2026, Zhipu listed on the Hong Kong Stock Exchange under the legal name Knowledge Atlas Technology (知识图谱科技). The IPO raised approximately HK$4.35 billion (roughly $558 million) at a share price of HK$116.20, valuing the company at about $6.5 billion. On its first day of trading, shares surged as high as HK$130 (an 11.88% gain), pushing the market capitalization above HK$57 billion ($7.4 billion). The Hong Kong public offering was oversubscribed 1,159 times, while the international offering saw 15.28 times oversubscription. Zhipu became both the first Chinese AI software company to go public and the world's first publicly traded company focused specifically on AGI foundation models.
In February 2026, Zhipu released GLM-5, its flagship model featuring approximately 745 billion parameters in a Mixture of Experts (MoE) architecture with 256 experts and 44 billion active parameters per inference token. GLM-5 uses DeepSeek Sparse Attention for efficient long-context handling and supports a 200K-token context window. The model was trained entirely on 100,000 domestically produced Huawei Ascend 910B chips using the MindSpore framework, a notable milestone for Chinese AI self-sufficiency given ongoing U.S. semiconductor export restrictions on NVIDIA GPUs. GLM-5 was released under the MIT license.
| Role | Name | Background |
|---|---|---|
| Chief Scientist and Co-founder | Tang Jie (唐杰) | Professor at Tsinghua University, ACM/IEEE/AAAI Fellow, Director of the Foundation Model Research Center at Tsinghua's AI Institute, vice director of BAAI, architect of the GLM model design |
| Co-founder | Li Juanzi (李涓子) | Professor at Tsinghua University's Department of Computer Science, co-lead of the Knowledge Engineering Group |
| Chairman and Co-founder | Liu Debing (刘德兵) | Former Technicolor (China) executive, oversees corporate governance and state-level alignment |
| CEO | Zhang Peng (张鹏) | Tsinghua University Computer Science alumnus, leads commercial strategy |
| President | Wang Shaolan (王绍兰) | Holds a doctorate, recognized as a Tsinghua Innovation Leader |
Tang Jie and Liu Debing serve as the effective controlling shareholders. Together with Li Juanzi, Zhang Peng, and other aligned stakeholders, the founding group controls approximately 36.97% of voting rights. Tang Jie directly owns 7.41% of the company.
The GLM (General Language Model) series is Zhipu's core technology. Unlike the purely autoregressive approach used by GPT models or the masked language modeling of BERT, GLM uses autoregressive blank infilling: the model learns to fill in randomly blanked-out spans of text in an autoregressive manner. This hybrid approach enables GLM to handle both natural language understanding and generation tasks within a single framework, providing versatility across downstream applications.
| Model | Release Date | Parameters | Key Features |
|---|---|---|---|
| GLM-130B | August 2022 | 130B | Bilingual (English/Chinese), open-weight, ICLR 2023 paper |
| ChatGLM-6B | March 2023 | 6.2B | Open-source, consumer-GPU friendly (6 GB INT4), optimized for Chinese dialogue |
| ChatGLM2-6B | June 2023 | 6.2B | Extended context (32K), 42% faster inference |
| ChatGLM3-6B | October 2023 | 6.2B | Native function calling, code interpreter, agent workflows |
| GLM-4 | January 2024 | Not disclosed (proprietary) | 128K context, 60% improvement over GLM-3, multilingual (26 languages) |
| GLM-4-9B | June 2024 | 9B | Open-source, web browsing, code execution, 128K context |
| GLM-4V-9B | June 2024 | 9B | Open-source multimodal vision-language model |
| GLM-4-Plus | August 2024 | Not disclosed | Proprietary, GPT-4o-comparable performance |
| GLM-4-32B-0414 | April 2025 | 32B | Open-source (MIT), strong reasoning, trained on 15 TB of data |
| GLM-Z1-32B-0414 | April 2025 | 32B | Reasoning model, comparable to DeepSeek-R1 on select tasks |
| GLM-4.5 | July 2025 | Not disclosed | Proprietary flagship, drove 10x usage growth |
| GLM-4.6V | December 2025 | 106B / 9B (Flash) | Vision-language, native function calling, 128K context |
| GLM-4.7 | December 2025 | Not disclosed | Open-source, optimized for coding and development workflows |
| GLM-5 | February 2026 | 745B (MoE, 44B active) | 200K context, trained on Huawei Ascend chips, MIT license |
Zhipu has maintained a strong commitment to open-source releases throughout its history. The ChatGLM-6B series was among the most popular open-source Chinese language models in 2023 and helped establish a large ecosystem of fine-tuned derivatives. Beginning with the GLM-4-0414 series in April 2025, Zhipu adopted the MIT license for its open-weight releases, allowing unrestricted commercial use and redistribution. This was a deliberate contrast to more restrictive licenses used by some competitors. Models are hosted on Hugging Face under the THUDM (Tsinghua University Data Mining Group) and later zai-org organization pages, and are also available through GitHub repositories.
The CogView series represents Zhipu's text-to-image generation technology, developed in collaboration with Tsinghua University's KEG and BAAI.
| Model | Year | Details |
|---|---|---|
| CogView | 2021 | 4B-parameter Transformer with VQ-VAE tokenizer, NeurIPS 2021 paper, outperformed prior GAN-based models on MS COCO |
| CogView2 | 2022 | Hierarchical Transformer, 10x faster than CogView, competitive with DALL-E 2 |
| CogView3 | 2024 | Cascaded diffusion model framework for higher quality generation |
| CogView3-Plus | 2024 | Based on Diffusion Transformer (DiT) architecture |
| CogView4 | 2025 | 6B parameters, native Chinese text input and Chinese text-to-image support |
In March 2025, Zhipu released CogKit, an open-source fine-tuning and inference framework that supports both the CogView4 and CogVideoX model families, allowing developers to customize these models for specific use cases.
CogVideo, and its successor CogVideoX, power Zhipu's video generation capabilities. The consumer-facing product derived from this technology is called Qingying (清影), which translates roughly to "clear reflection." Qingying was launched in July 2024 as a text-to-video generator available to the public.
| Model | Year | Details |
|---|---|---|
| CogVideo | 2022 | Early text-to-video Transformer model, ICLR 2023 paper |
| CogVideoX-2B | 2024 | 2B parameters, open-source text-to-video, same lineage as Qingying |
| CogVideoX-5B | 2024 | 5B parameters, enhanced video duration, spatial resolution, and semantic consistency |
| CogVideoX-5B-I2V | September 2024 | Image-to-video generation, open-source |
| CogVideoX v1.5 | Late 2024 | 768P resolution, 5-second and 10-second video lengths, 16 frames per second |
CogVLM is a family of vision-language models developed by Zhipu and Tsinghua. CogVLM-17B was open-sourced in October 2023 alongside ChatGLM3 and demonstrated strong performance on visual question answering and image understanding tasks. The CogVLM line later evolved into the GLM-4V series of integrated multimodal models.
Zhipu Qingyan is the company's consumer-facing chatbot application, comparable to ChatGPT in function. The app features text conversation, code interpretation, video call interaction, and multimodal input. It provides access to the latest GLM models along with specialized tools for image generation, web browsing, and data analysis. By late 2025, Zhipu Qingyan had accumulated over 25 million registered users, making it one of the most widely used AI assistant apps in China.
The Zhipu Open Platform, accessible at bigmodel.cn (and internationally at open.bigmodel.cn), provides a Model-as-a-Service (MaaS) API for developers and enterprises. The platform offers access to the full GLM model family, including language models, vision models, image generation (CogView), and video generation (CogVideoX). It supports model fine-tuning, evaluation tools, intelligent agent creation, and a one-stop modular toolkit for building applications with large models. Fine-tuning can be completed in as little as ten minutes for lightweight tasks. Pricing follows a pay-as-you-go model; as of mid-2025, GLM-4.x API access cost approximately $0.11 per million input tokens. A Batch API option is also available at 50% of standard cost for large-scale data processing. In June 2024, Zhipu slashed GLM-4 API prices by over 50%, joining a broader price war among Chinese AI providers.
AutoGLM is Zhipu's AI agent platform designed for autonomous device operation. The system interprets on-screen content and performs human-like interactions (taps, swipes, text entry) to complete multi-step tasks such as food delivery orders, flight bookings, and cross-app workflows. First introduced in November 2024, AutoGLM was later open-sourced in December 2025, allowing developers to build on its device-operation capabilities.
Zhipu provides localized deployment solutions for enterprise customers, which accounted for 84.8% of the company's first-half 2025 revenue. Cloud-based deployment made up the remaining 15.2%. Strategic partnerships include a collaboration with Huawei on "AI-in-a-box" solutions for on-premises data processing and integration work with Qualcomm for on-device AI deployment in mobile and automotive applications. The company also offers government-focused AI solutions, serving both domestic Chinese government customers and exploring international government partnerships.
Zhipu has raised approximately $1.5 billion in total funding across 12 rounds since its founding, including 6 early-stage and 6 late-stage rounds.
| Period | Amount | Notable Investors | Valuation |
|---|---|---|---|
| Early rounds (2019-2022) | Undisclosed | Various Chinese VCs | N/A |
| October 2023 | ~$342M | Alibaba, Tencent, Meituan, Xiaomi, HongShan | ~$2.8B |
| May 2024 | ~$400M | Prosperity7 Ventures (Saudi Aramco), Ant Group | ~$3.0B |
| December 2024 (Series D) | ~$420M | Multiple investors | ~$4.0B |
| March 2025 | ~$257M | Chengdu, Zhuhai, and Hangzhou government-backed funds | N/A |
| July 2025 | ~$140M | Shanghai state funds | N/A |
| January 2026 (IPO) | ~$558M | Public market | ~$6.5B |
The presence of government-backed funds in the 2025 rounds reflects broader Chinese state support for domestic AI development and self-sufficiency.
| Year | Revenue (RMB) | Net Loss (RMB) |
|---|---|---|
| 2022 | 57 million | 144 million |
| 2023 | 125 million | 788 million |
| 2024 | 312 million | 2.96 billion |
| H1 2025 | 191 million | 2.36 billion |
Revenue grew at a compound annual growth rate of approximately 130% from 2022 to 2024. However, losses grew substantially alongside revenue, driven primarily by R&D spending on model training and infrastructure. R&D expenses increased from RMB 84 million in 2022 to approximately RMB 5.29 billion in 2023. The majority of revenue comes from localized (on-premises) deployment services for enterprise customers, with cloud-based API services making up a smaller share. This pattern is typical of Chinese enterprise AI companies, where large clients often require private deployment for data security and compliance reasons.
Zhipu is recognized as one of China's "Six Little Tigers" (六小虎) of AI, a term coined by Chinese media outlet Caixin in late 2024 to describe the country's most prominent generative AI startups. The six companies are:
| Company | Focus | Founded |
|---|---|---|
| Zhipu AI (Z.ai) | General-purpose LLMs, multimodal AI, AI agents | 2019 |
| Baichuan Intelligence | Large language models | 2023 |
| MiniMax | Multimodal AI, consumer applications | 2021 |
| Moonshot AI (Dark Side of the Moon) | Long-context language models (Kimi) | 2023 |
| 01.AI (Zero One Everything) | Open-source LLMs (Yi series) | 2023 |
| StepFun (Leap Star) | Multimodal foundation models | 2023 |
Among these companies, Zhipu stands out as the earliest founded, the largest by headcount (approximately 800 employees as of late 2024, with 60-70% in R&D roles), and the first to pursue a public offering. Each of the Six Little Tigers is valued at several billion dollars. The company became the first of the group to complete an IPO when it listed in Hong Kong in January 2026, followed shortly by MiniMax.
In the Chinese AI market, Zhipu competes with both the Six Little Tigers and the AI divisions of major technology companies such as Baidu (ERNIE), Alibaba (Qwen), ByteDance (Doubao), and Tencent (Hunyuan). The company also faces competition from DeepSeek, which gained international attention in early 2025 with its cost-efficient reasoning models and accelerated AI investment across China.
Internationally, Zhipu positions itself as a competitor to OpenAI, Anthropic, and Google DeepMind. The company's stated mission is to pursue artificial general intelligence (AGI) with "super cognitive intelligence beyond human level."
Zhipu has differentiated itself through several strategic choices: a strong open-source commitment with permissive MIT licensing, dual focus on both consumer (Zhipu Qingyan) and enterprise (MaaS platform) markets, deep academic roots at Tsinghua University providing a steady pipeline of research talent, and a willingness to train frontier models on domestically produced hardware (Huawei Ascend chips) rather than relying solely on restricted NVIDIA GPUs. This last factor has become increasingly important given ongoing U.S. export controls that limit Chinese companies' access to advanced NVIDIA chips like the A100, H100, and H200.
Beyond its commercial products, Zhipu and its affiliated researchers have made notable academic contributions:
Key papers from the team include "GLM-130B: An Open Bilingual Pre-trained Model" (ICLR 2023), "CogView: Mastering Text-to-Image Generation via Transformers" (NeurIPS 2021), "CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers" (ICLR 2023), and "ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools" (2024).