Z.ai
Last reviewed
Sources
43 citations
Review status
Source-backed
Revision
v7 · 7,054 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
43 citations
Review status
Source-backed
Revision
v7 · 7,054 words
Add missing citations, update stale details, or suggest a clearer explanation.
| Z.ai | |
|---|---|
| 智谱 | |
 | |
| Type | Public company (HKEX: 2513) |
| Industry | Artificial intelligence |
| Founded | June 11, 2019 |
| Founders | Tang Jie, Li Juanzi |
| Headquarters | Haidian District, Beijing, China |
| Key people | Tang Jie (Co-founder) Li Juanzi (Co-founder) Liu Debing (Chairman) Zhang Peng (CEO) |
| Products | GLM (General Language Model) family, ChatGLM, GLM-4, GLM-4.5, GLM-4.6, GLM-4.7, GLM-5, GLM-5.1, GLM-5.2, CodeGeeX, AutoGLM, Ying, CogVLM, CogView, CogVideoX, GLM-Image, Zhipu Qingyan, MaaS Platform |
| Revenue | RMB 724 million (2025), up 131.9% YoY |
| Listed name | Knowledge Atlas Technology Joint Stock Co., Ltd. |
| IPO | Hong Kong Stock Exchange, January 8, 2026 |
| Employees | ~2,000 (2025) |
| Website | z.ai / zhipuai.cn |
Z.ai is the international brand of the Chinese artificial intelligence company Zhipu AI (智谱AI), a 2019 spinout from Tsinghua University that builds the open-weight General Language Model (GLM) family and became, in January 2026, the first foundation-model company in the world to complete an initial public offering, listing on the Hong Kong Stock Exchange under the ticker 2513.[19][20] The legal Chinese operating entity is Beijing Zhipu Huazhang Technology Co., Ltd. (北京智谱华章科技有限公司), with the listed parent named Knowledge Atlas Technology Joint Stock Co., Ltd.; the company adopted the global "Z.ai" identity in July 2025 alongside the release of GLM-4.5.[11][13] Z.ai is consistently grouped among China's so-called "AI Tigers" (also called the "Six Little Tigers") alongside Moonshot AI, Baichuan Intelligence, MiniMax, StepFun, and 01.AI.[33]
Z.ai develops the General Language Model (GLM) series, which includes GLM-130B, the ChatGLM family of open-weight chat models, the multimodal CogVLM and CogVideoX models, the CodeGeeX code generation suite, the AutoGLM mobile agent platform, and the flagship GLM-4, GLM-4.5, GLM-4.6, GLM-4.7, GLM-5, GLM-5.1, and GLM-5.2 foundation models. Industry analyses through 2025 and 2026 regularly placed the GLM family among the top open-weight model lines globally, along with offerings from DeepSeek, Alibaba's Qwen team, and Moonshot AI's Kimi.[15] By April 2026 the company reported serving more than four million small and medium-sized businesses and developers across 218 countries and regions, and its 2025 revenue reached RMB 724 million, up 131.9 percent year over year.[27][32]
Yes. Z.ai and Zhipu AI are the same company under two brand names. "Zhipu AI" (智谱AI) is the original and still-current name used in the Chinese market; "Z.ai" is the international brand the company adopted on July 28, 2025, concurrent with the GLM-4.5 release, to present a cleaner global identity.[11][13] The underlying legal entity in China is Beijing Zhipu Huazhang Technology Co., Ltd., and the Hong Kong-listed parent is Knowledge Atlas Technology Joint Stock Co., Ltd., a literal English rendering of the parent's revised Chinese name 知识图谱科技 ("knowledge graph technology").[22] In June 2026, when the board approved a second listing on Shanghai's STAR Market, the company also proposed changing its English corporate name to "Z.AI Co., Ltd." to align the legal name with the brand.[37][38] This wiki maintains a companion page at Zhipu AI covering the same organization.
Zhipu AI was founded on June 11, 2019 by Tsinghua University professors Tang Jie and Li Juanzi at the Tsinghua University Science Park in Beijing's Zhongguancun district.[1] The company originated from Tsinghua's Knowledge Engineering Group, a research lab that Li Juanzi continues to direct and that has also incubated other AI startups including Moonshot AI and DeepLang. The founders faced early challenges securing investment in a market that had not yet adopted the foundation model paradigm, and the administrative commission of Zhongguancun Science Park provided the team with three months of rent-free office space to help kickstart development.
The firm initially focused on knowledge graph technology, drawing on a long lineage of academic work from Tang Jie's group, including the AMiner academic search platform that he had been building since the mid-2000s. By late 2020 the team pivoted to large model AI, betting that a new generation of bidirectional pre-trained transformers would dominate Chinese-language natural language processing. The pivot was contemporaneous with the publication of the original GLM (General Language Model) algorithm paper at the Association for Computational Linguistics conference in 2022, which introduced an "autoregressive blank infilling" pre-training objective that unified bidirectional understanding and left-to-right generation in a single architecture.
In late 2020 Zhipu developed the GLM pre-training architecture. By 2021 the team had completed training of the GLM-10B model with tens of billions of parameters. In 2022 the company developed and open-sourced GLM-130B, a 130-billion-parameter bilingual pre-training model supporting Chinese and English. GLM-130B was pre-trained on 400 billion tokens using a cluster of 96 NVIDIA DGX A100 (8 by 40 GB) nodes between May 6 and July 3, 2022, and was accepted to ICLR 2023.[2] Public benchmarks at the time indicated that GLM-130B matched or surpassed GPT-3 175B (davinci) on a wide range of English tasks and significantly outperformed Baidu's ERNIE TITAN 3.0 260B across major Chinese benchmarks.[2] The release made GLM-130B one of the most capable openly available LLMs in the world during the brief window between BLOOM and Meta's LLaMA, and put Zhipu on the map as a serious foundation model lab.
In 2023 Zhipu turned the GLM architecture into a consumer-facing chat assistant. ChatGLM-6B, an open-weight 6-billion-parameter chat model released on March 14, 2023, was the first widely downloadable Chinese instruction-tuned LLM with strong performance and quickly became one of the most downloaded models of the year on Hugging Face.[6] It supported local inference on a single consumer GPU and was rapidly fine-tuned by hobbyists, university labs, and enterprises across China. A larger ChatGLM model with hundreds of billions of parameters was offered as a paid API through the Zhipu Qingyan (智谱清言) consumer chatbot.
ChatGLM2-6B followed on June 25, 2023. Using a hybrid GLM objective and 1.4 trillion bilingual training tokens, it achieved substantial gains on MMLU, CEval, GSM8K, and BBH versus its predecessor, while extending the base context window from 2K to 32K tokens through FlashAttention. The release also introduced Multi-Query Attention, raising inference speed roughly 42 percent.
ChatGLM3-6B arrived on October 27, 2023 with a redesigned prompt format that natively supported function calling, code interpreter use, and complex agent tasks. The ChatGLM3-6B-Base variant achieved the strongest performance on standard reasoning, math, and code benchmarks among pre-trained models under 10 billion parameters at the time of release. By the end of 2023 the ChatGLM family had been downloaded more than 30 million times globally and was the dominant open Chinese chat model.[6]
During the same period, Zhipu and Tsinghua KEG also released CodeGeeX, an open multilingual code generation model trained on more than 850 billion tokens across 20-plus programming languages on a cluster of 1,536 Huawei Ascend 910 AI processors. CodeGeeX was published at KDD 2023, distributed as a free Visual Studio Code and JetBrains plug-in, and supported function-level completion, code translation, and code explanation.[7]
In January 2024 the company unveiled GLM-4, its next-generation flagship foundation model. GLM-4 introduced a 128K-token context window, native multimodal understanding spanning text, image, and video, and the "All Tools" agentic feature that allowed the model to autonomously use a web browser, a code interpreter, and image generation tools.[6] Internal benchmarks placed GLM-4 within striking distance of GPT-4 on Chinese reasoning, English MMLU, GSM8K, and HumanEval, and the model became the backbone of both the Zhipu Qingyan consumer app and the company's enterprise MaaS (Model-as-a-Service) platform.
In March 2024 CEO Zhang Peng publicly announced that Zhipu was developing a Sora-style text-to-video model as a step toward artificial general intelligence. The company debuted that system, branded "Ying" (清影, Qingying), in July 2024 with the ability to generate six-second video clips, and later open-sourced the underlying technology as CogVideoX, a DiT (diffusion transformer) family using a 3D causal variational autoencoder. CogVideoX-5B was released under an Apache 2.0 license, with CogVideoX v1.5 in late 2024 extending generation to 5-second and 10-second clips at 768p resolution and 16 frames per second.
Alongside model launches, Zhipu pursued a near-continuous fundraising campaign in 2024. A consortium of Alibaba, Tencent, Meituan, Ant Group, Xiaomi, and HongShan (formerly Sequoia China) had contributed roughly 2.5 billion yuan in October 2023; in May 2024, Saudi Aramco's Prosperity7 Ventures led a roughly 400 million dollar round that pushed the company's valuation to about 3 billion dollars.[29] In December 2024 a Series D worth approximately 411 million dollars valued the company at around 2.8 billion dollars, with Alibaba and other corporates returning and Beijing and Shenzhen municipal funds taking minority stakes. The Beijing AI Industry Investment Fund and Zhuhai's state-owned Huafa Group emerged as among Zhipu's most influential backers.[29]
In October 2024 the company introduced AutoGLM, an autonomous agent that could operate smartphones via voice commands. AutoGLM was demonstrated executing multi-step workflows across more than fifty Chinese applications, including WeChat, Taobao, Meituan, Ctrip, Douyin, Xiaohongshu, and the 12306 train ticketing platform. A subsequent upgrade extended task length to more than 54 steps and added integration with Alibaba Cloud for smartphone deployment.
On January 16, 2025 the United States Department of Commerce's Bureau of Industry and Security added Zhipu's corporate parent, Beijing Zhipu Huazhang Technology Co., Ltd., and ten affiliated Zhipu entities to the U.S. Entity List under Section 744.11 of the Export Administration Regulations, stating that the listed entities "advance the People's Republic of China's military modernization through the development and integration of advanced artificial intelligence research."[8][15] The designation made Zhipu the first Chinese large model company added to the list, and it imposed a license requirement, with a presumption of denial, for exports, reexports, and transfers of items subject to the EAR.[31] Zhipu publicly rejected the allegations as lacking factual basis and stated that the listing would not materially impact its operations because the company did not rely on U.S. large-model technology.[9]
In response to the export controls, Zhipu accelerated its transition toward domestic compute. Engineers ported training and inference workloads to Huawei Ascend processors using the MindSpore framework, and signed broader partnerships with Huawei, Cambricon, and several state-backed cloud providers. The company also continued to raise money. A 257 million dollar round in March 2025 was led by government-backed funds tied to Hangzhou, Zhuhai, and Chengdu, while a roughly 140 million dollar strategic round in July 2025 brought in Shanghai state funds including Pudong Venture Capital and Zhangjiang Group.[17][19][29] By mid-2025 Zhipu had received capital from the AI industry funds of Beijing, Shanghai, Shenzhen, Hangzhou, and Zhuhai, illustrating how local governments across China were competing to back national AI champions.
In April 2025 Zhipu filed for IPO counseling with China International Capital Corporation, becoming the first of the AI Tigers to publicly disclose plans for a stock market listing. The company restructured from a foreign-invested limited liability company into a foreign-invested joint-stock company in preparation. Internally the entity adopted the trading name Knowledge Atlas Technology Joint Stock Co., Ltd., a literal English rendering of the parent company's revised Chinese name. Around the same time Zhipu open-sourced the GLM-4-0414 series under the MIT license, including a 32-billion-parameter base model and the GLM-Z1-32B-0414 reasoning model that matched aspects of DeepSeek-R1 (a 671-billion-parameter model) on select tasks despite its far smaller size.[8]
On July 28, 2025 Zhipu announced its global rebrand to Z.ai concurrent with the release of GLM-4.5 and GLM-4.5 Air.[11][13] GLM-4.5 is a 355-billion-parameter Mixture-of-Experts foundation model with 32 billion active parameters per forward pass.[12] The accompanying technical report on arXiv, titled "GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models" (arXiv:2508.06471), describes the system as an "open-source Mixture-of-Experts (MoE) large language model with 355B total parameters and 32B activated parameters" trained through "multi-stage training on 23T tokens and comprehensive post-training with expert model iteration and reinforcement learning."[5] On the report's evaluation suite GLM-4.5 scored 70.1 percent on TAU-bench, 91.0 percent on AIME 2024, and 64.2 percent on SWE-bench Verified, and the authors stated that "with much fewer parameters than several competitors, GLM-4.5 ranks 3rd overall among all evaluated models and 2nd on agentic benchmarks."[5] GLM-4.5-Air offered a leaner 106-billion-parameter variant (12 billion active) for edge deployments. Both were released under the permissive MIT license with weights published on Hugging Face, and Z.ai listed the API at roughly 0.11 dollars per million input tokens and 0.28 dollars per million output tokens, undercutting DeepSeek and the major U.S. labs.[10][12]
In August 2025 Z.ai followed up with GLM-4.5V, a vision-language model based on GLM-4.5-Air that reported state-of-the-art results across roughly 42 public vision-language benchmarks among open models of its size.[22] AutoGLM 2.0 launched the same month, positioned as the world's first cross-application mobile agent.
GLM-4.6, released on September 30, 2025, is an open-weight (MIT-licensed) model focused on agentic coding, with a 200,000-token context window, up from 128K in GLM-4.5.[14][23] On Z.ai's CC-Bench evaluation, where human raters ran multi-turn coding tasks in isolated Docker environments, GLM-4.6 reached near parity with Anthropic's Claude Sonnet 4 (a 48.6 percent win rate) while using about 15 percent fewer tokens per task than GLM-4.5, and the weights were again published openly on Hugging Face for local deployment.[14][23] The release also marked the first public deployment of FP8 and Int4 quantization on Cambricon chips, broadening the model's deployable hardware footprint inside China. In September 2025 the company also offered explicit migration guides for users of Claude to switch to GLM-4.5 and GLM-4.6, capitalizing on Anthropic's tightening usage caps and access controls in some regions.[14]
December 2025 brought two major releases. On December 9, 2025 Zhipu open-sourced AutoGLM as a full mobile-agent framework powered by GLM-4.5 and GLM-4.5V, the first such open release in the industry.[30] On December 22, 2025 the company shipped GLM-4.7, a coding-centric 355B/32B-active MoE model with a 200K context window and up to 128K-token outputs.[18] GLM-4.7 scored 73.8 percent on SWE-bench Verified, the highest result among open-source models at the time, and 84.9 percent on LiveCodeBench, and it ranked first among open-source and Chinese models on the Code Arena blind-testing leaderboard.[15][18][28] Z.ai also released GLM-4.6V, a 106-billion-parameter open-weight vision-language model with native tool calling and a 128K context, plus a lightweight 9-billion-parameter GLM-4.6V-Flash variant.[17] Both were optimized for GUI agents and frontend automation.[16]
On January 8, 2026 Z.ai's parent company Knowledge Atlas Technology Joint Stock Co., Ltd. listed on the Hong Kong Stock Exchange under the ticker 2513.[20] The company issued 37.42 million H-shares at HK$116.20 per share, raising roughly HK$4.35 billion (about 558 million U.S. dollars).[21] Knowledge Atlas opened at HK$120.00, surged as high as HK$130 (an 11.88 percent gain), and pushed its market capitalization above HK$57 billion on the first day, becoming the first of China's foundation-model AI Tigers to go public and the first publicly listed company focused specifically on AGI foundation models.[19][20] The Hong Kong public offering was oversubscribed 1,159 times and the international tranche 15.28 times, and the prospectus indicated that roughly 70 percent of IPO proceeds would be deployed into R&D for general-purpose large models, with the remainder allocated to commercialization and overseas expansion.[6][22]
On January 14, 2026 Z.ai unveiled GLM-Image, a multimodal model that was the first state-of-the-art Chinese AI model trained end-to-end on Huawei Ascend hardware (the Atlas 800T A2 device) using MindSpore.[24] The release showcased Zhipu's ability to ship frontier models without any Nvidia silicon and validated China's domestic semiconductor stack for advanced AI workloads.[23] To achieve this, Z.ai engineers developed dynamic graph multi-level pipelined deployment, custom fusion operators tuned for Ascend's architecture, and multi-stream parallelism that overlapped communication with computation during distributed training.
On February 11, 2026 Z.ai released GLM-5, a 744-billion-parameter Mixture-of-Experts model with 40 billion active parameters per token, accompanied by a technical report titled "GLM-5: from Vibe Coding to Agentic Engineering" describing pre-training on 28.5 trillion tokens and an asynchronous reinforcement learning post-training infrastructure.[25][30] GLM-5 reached 77.8 percent on SWE-bench Verified, the highest result among open-weight models at the time and within roughly three points of GPT-5.2 and Claude Opus 4.5, and it ranked first among open models on the LMArena text leaderboard.[25] The full GLM-5 training run again used Huawei Ascend hardware with zero Nvidia dependency, and the weights shipped under the MIT license. Z.ai priced the GLM-5 API at $1.00 per million input tokens and $3.20 per million output tokens, roughly five times cheaper on input and eight times cheaper on output than Claude Opus.[25] Zhipu's shares rose 34 percent on the day of the release, and by mid-February 2026 the stock had gained more than 500 percent from its IPO price, lifting the market capitalization past 40 billion dollars.[25]
A brief pullback came later that month: after demand for GLM-5 exceeded inference capacity, Z.ai temporarily restricted new sign-ups to its API and the stock declined for a period before stabilizing. Compute shortages remained a defining constraint of the Chinese AI sector through 2026, with state-led investment in additional Ascend and Cambricon production attempting to close the gap.
On April 8, 2026 Z.ai released GLM-5.1, an open-weight 744-billion-parameter agentic foundation model that the company described as its most intelligent flagship to date, capable of working continuously on a single task for more than eight hours.[26] Zhipu said GLM-5.1 ranked first among open-source and Chinese models, and third overall, on the combined average of SWE-bench Pro, Terminal-Bench, and NL2Repo.[26] The release accompanied a roughly 10 percent average increase in cloud API prices, the company's second price rise of 2026, and the shares climbed nearly 19 percent intraday on the announcement.[26] The company's first post-IPO annual report, issued at the end of March 2026, disclosed that 2025 revenue reached 724 million yuan, up 131.9 percent year over year, while the net loss widened to RMB 4.72 billion on heavy R&D spending.[28][32] Open platform and API revenue rose 292.6 percent to RMB 190 million, the enterprise intelligent-agent business grew 248.8 percent to RMB 166 million, and enterprise-deployed model revenue grew 70.5 percent to RMB 366 million; the shares rose about 32 percent to HK$915 on April 1, 2026 after the results.[28][32] By April 2026 the company served more than four million SMEs and developers across 218 countries and regions.[27]
The steady cadence of releases and surging API demand drove an extended stock rally. On May 28, 2026 Z.ai shares briefly touched HK$1,993, valuing the company at more than HK$880 billion (about 112 billion dollars), nearly 1,600 percent above the IPO price, before partially reversing.[36] On June 1, 2026 the board unanimously approved a plan for a second listing on Shanghai's STAR Market, seeking to raise up to RMB 15 billion (about 2.1 billion dollars) through an A-share issuance and creating a dual A+H share structure; the resolution also proposed renaming the corporate entity "Z.AI Co., Ltd." to match the brand.[37][38]
On June 13, 2026 Z.ai announced GLM-5.2, an upgrade to the GLM-5 series that quadrupled the context window to 1 million tokens while retaining the 744-billion-parameter MoE architecture, releasing the weights under the MIT license on Hugging Face and ModelScope.[41][42] In independent benchmark coverage, GLM-5.2 reported 62.1 on SWE-bench Pro, ahead of OpenAI's GPT-5.5 (58.6) and the predecessor GLM-5.1 (58.4), and 81.0 on Terminal-Bench 2.1, within about one point of Anthropic's Claude Opus 4.8 on long-horizon coding tasks while undercutting comparable proprietary APIs on price.[42][43] In its release notes the company cautioned that "a 1M context is easy to claim, but much harder to keep reliable under real engineering pressure," framing reliability rather than raw context length as the engineering challenge.[42] News of the launch sent the shares up as much as 48 percent intraday on June 15, 2026 before they closed up 32.8 percent at HK$1,457.[40]
Tang Jie is co-founder, controlling shareholder, and the central technical figure of Z.ai. A professor in Tsinghua University's Department of Computer Science and Technology and director of the Foundation Model Research Center at Tsinghua's AI Institute, Tang is a Fellow of the IEEE, ACM, and AAAI and serves as vice director of the Beijing Academy of Artificial Intelligence (BAAI), the non-profit research institute that produced the Wu Dao 2.0 model in 2021. Tang directly owns about 7.41 percent of the company and controls a larger share of voting rights through shareholding platforms.[22] He has personally mentored a generation of Chinese AI researchers, including founders of rival labs such as Moonshot AI.
Li Juanzi is co-founder and a professor in Tsinghua's Department of Computer Science and Technology, where she continues to direct the Knowledge Engineering Group. Beyond Zhipu, KEG has incubated DeepLang and Moonshot AI among other startups, making it one of the most prolific lab origins for Chinese AI ventures.
Liu Debing serves as chairman and co-controlling shareholder, a former Technicolor (China) executive who oversees corporate governance and state-level alignment. With Tang Jie and other aligned stakeholders he is part of a control group that collectively holds about 36.97 percent of voting rights, since the company has no single majority shareholder.
Zhang Peng is the chief executive officer. He earned a degree in computer science from Tsinghua University and has been the public face of Zhipu's strategic announcements, including the rebrand to Z.ai, the GLM-4 and GLM-4.5 releases, and the company's IPO. After the 2025 annual results, Zhang said Zhipu aspires to become "China's Anthropic," with API revenue serving as the foundation for sustained, long-term expansion.[32]
| Role | Name | Voting interest (approx.) |
|---|---|---|
| Co-founder, controlling shareholder | Tang Jie | 7.41% direct; larger via platforms |
| Chairman, co-controlling shareholder | Liu Debing | aligned with Tang |
| CEO, co-controlling group | Zhang Peng | aligned with Tang and Liu |
| Aggregate concert party | Tang, Liu, Zhang, allies | ~36.97% |
The company has no single controlling shareholder, with a concert party of co-founders, executives, and early backers exercising collective control through an acting-in-concert agreement disclosed in the IPO prospectus.[22]
Z.ai's product portfolio spans foundation models, multimodal models, code tools, autonomous agents, and a consumer chat app. The GLM-4 family and its successors anchor the lineup, with ChatGLM serving as the on-ramp for open-weight research deployments.[4]
The General Language Model series is the company's flagship line of pre-trained models, descended from the GLM algorithm published at ACL 2022 and the GLM-130B paper accepted at ICLR 2023.
| Model | Release | Parameters | Notable features |
|---|---|---|---|
| GLM-10B | 2021 | ~10B | Early bilingual model, internal research |
| GLM-130B | August 2022 | 130B | Open bilingual model, matched GPT-3 175B, ICLR 2023 |
| ChatGLM-6B | March 14, 2023 | 6B | First open Chinese chat model, runs on consumer GPU |
| ChatGLM2-6B | June 25, 2023 | 6B | 32K context, Multi-Query Attention, +42% inference speed |
| ChatGLM3-6B | October 27, 2023 | 6B | Function calling, code interpreter, agent prompts |
| GLM-4 | January 2024 | Undisclosed | 128K context, multimodal, All Tools agent |
| GLM-4-Plus | August 2024 | Undisclosed | Enhanced reasoning and code |
| GLM-4-Voice | October 2024 | Undisclosed | End-to-end speech LLM |
| GLM-4-0414 / GLM-Z1-32B-0414 | April 2025 | 32B | Open weights (MIT), reasoning model near DeepSeek-R1 on select tasks |
| GLM-4.5 / GLM-4.5 Air | July 28, 2025 | 355B (32B active) / 106B (12B active) | MoE, MIT open weights, 23T-token training, 64.2% SWE-bench Verified |
| GLM-4.5V | August 2025 | 106B class | SOTA on ~42 vision-language benchmarks |
| GLM-4.6 | September 30, 2025 | 355B (32B active) | 200K context, Cambricon FP8/Int4 support, agentic coding |
| GLM-4.7 | December 22, 2025 | 355B (32B active) | 73.8% SWE-bench Verified, persistent thinking blocks |
| GLM-4.6V / GLM-4.6V-Flash | December 9, 2025 | 106B / 9B | Native tool calling, 128K context, GUI agents |
| GLM-Image | January 14, 2026 | Not disclosed | First SOTA Chinese model fully trained on Huawei Ascend |
| GLM-5 | February 11, 2026 | 744B (40B active) | 77.8% SWE-bench Verified, trained on Ascend, 200K context |
| GLM-5.1 | April 8, 2026 | 744B (40B active) | SOTA on SWE-Bench Pro combination, 8-hour autonomous execution |
| GLM-5.2 | June 13, 2026 | 744B (40B active) | 1M-token context, 62.1 SWE-bench Pro, MIT open weights |
Key architectural choices across the modern GLM family include a Mixture-of-Experts core with sigmoid gates, Grouped-Query Attention with partial Rotary Positional Embeddings, QK-norm, MoE layers that double as multi-token prediction layers to enable speculative decoding, and an unusually deep configuration that uses roughly 2.5 times more attention heads than typical for similar hidden dimensions. Z.ai positions the models as agentic, reasoning, and coding foundation models, with a dual-mode operation that exposes a fast "non-thinking" path for low-latency queries and a deliberate "thinking" path for hard problems.[5]
Yes. Beginning with the GLM-4-0414 series in April 2025, every Z.ai flagship has shipped open weights under the permissive MIT license, which allows unrestricted commercial use, fine-tuning, and local deployment.[8] GLM-4.5, GLM-4.6, GLM-4.7, GLM-5, GLM-5.1, and GLM-5.2 were all released this way on Hugging Face (and, for the newest models, ModelScope), with technical reports published for the major releases.[5][41][42] This permissive licensing is a deliberate contrast to the more restrictive community licenses used by some competitors and is central to Z.ai's strategy of seeding a global ecosystem of GLM derivatives. The weights are hosted under the THUDM (Tsinghua University Data Mining) and later zai-org GitHub and Hugging Face organizations.[3]
AutoGLM is the company's autonomous agent platform, first released in October 2024 and open-sourced as a full framework on December 9, 2025. AutoGLM uses GLM-4.5 and GLM-4.5V (and later GLM-4.6 and GLM-4.6V) to perform on-device "phone use" tasks. Notable capabilities include:
AutoGLM 2.0 launched in August 2025 and was positioned as the world's first cross-app mobile agent. The open-source release in December 2025, branded Open-AutoGLM, made the framework available for community deployment, including the AutoGLM-Phone-9B model weights and local-only operation.[30]
CogVLM and CogView were Zhipu's earlier vision-language and text-to-image models. CogVideoX, released in 2024 under an Apache 2.0 license, is a DiT-based open-source text-to-video model using a 3D causal variational autoencoder. CogVideoX-5B and the later CogVideoX v1.5 family support generation of 5-second and 10-second clips at 768p and 16 frames per second, with image-to-video variants. The Ying (清影) text-to-video assistant, launched in July 2024, generated six-second clips as Z.ai's consumer-facing video product. GLM-4.5V and GLM-4.6V extended the multimodal lineage into reasoning-capable VLMs, with GLM-4.6V adding native tool calling that lets a model pass screenshots, documents, and images directly to downstream tools without going through a text bottleneck.[16]
CodeGeeX is the company's family of code generation models, originating from a 13-billion-parameter pre-trained model trained on more than 850 billion tokens across 20-plus programming languages on a cluster of 1,536 Huawei Ascend 910 processors. Released as a free Visual Studio Code and JetBrains extension, CodeGeeX supports code completion, function-level generation, code translation, and code explanation. Surveys cited in the KDD 2023 paper indicated that 83.4 percent of users reported improved programming efficiency.[7] By 2025 the CodeGeeX tooling had been integrated with the GLM-4, GLM-4.6, and GLM-4.7 models, and Z.ai launched a GLM Coding Plan that paired the GLM Coding API with popular agent harnesses including Claude Code, Cline, and Roo Code. At the GLM-4.7 launch in December 2025 the entry (Lite) tier was priced at about $3 per month, a figure widely contrasted in press coverage with U.S. coding subscriptions costing as much as $200 per month; the plan was later repriced upward as demand grew, with new-subscriber rates starting around $18 per month by 2026.[29]
Zhipu Qingyan (智谱清言) is the company's consumer-facing AI assistant in mainland China, exposing the GLM family for conversation, document analysis, image generation, video generation, and agent workflows. By late 2025 the app reported more than 25 million registered users, making it one of the most widely used AI assistant apps in China. Internationally, Z.ai operates the chat.z.ai web product, which became the consumer entry point for GLM-4.6 and successor models outside China.
Z.ai raised roughly 1.5 billion U.S. dollars across about a dozen funding rounds prior to its IPO. The investor base spans Chinese internet majors, prominent Chinese venture firms, Saudi sovereign capital, and a long list of state-backed municipal funds.
| Date | Round | Amount | Lead investors | Valuation |
|---|---|---|---|---|
| September 2022 | Series B | Undisclosed | Qiming Venture Partners | Undisclosed |
| October 2023 | Series B+ | ~$342 million (RMB 2.5B) | Alibaba, Tencent, Meituan, Ant Group, Xiaomi, HongShan | ~$2.8 billion |
| May 2024 | Series C | ~$400 million | Prosperity7 Ventures (Saudi Aramco) | ~$3 billion |
| December 2024 | Series D | ~$411 million | Alibaba, Tencent, BAII | ~$2.8 billion |
| March 2025 | State-backed round | ~$257 million | Hangzhou, Zhuhai, Chengdu government funds | N/A |
| July 2025 | Strategic | ~$140 million | Pudong Venture Capital, Zhangjiang Group (Shanghai state funds) | N/A |
| January 8, 2026 | IPO (HKEX: 2513) | ~$558 million (HK$4.35B) | Hong Kong public market | ~HK$57 billion mcap at debut |
Major pre-IPO investors include Alibaba Group, Tencent Holdings, Ant Group, Meituan, Xiaomi, HongShan (formerly Sequoia China), Legend Capital, Qiming Venture Partners, Prosperity7 Ventures, the Beijing AI Industry Investment Fund, Zhuhai Huafa Group, Hangzhou Industrial Investment Group, Pudong Venture Capital, and Zhangjiang Group. The breadth of state-backed and corporate Chinese investors reflects Z.ai's status as a politically and commercially strategic asset within China's AI ecosystem.[29] The Hong Kong IPO drew oversubscription of 1,159 times on the public tranche, and the stock subsequently traded many multiples above its offer price during a sustained post-listing rally that peaked near HK$1,993 in late May 2026.[19][36]
Z.ai monetizes through three primary channels. Enterprise model deployments, mostly delivered on customer premises or on dedicated cloud infrastructure, accounted for roughly half of 2025 revenue (RMB 366 million). The company's open platform and API business, including the MaaS service that gives developers metered access to GLM models, grew 292.6 percent in 2025 to RMB 190 million, with annualized recurring revenue from the platform reaching roughly RMB 1.7 billion by March 2026. The enterprise agent business, which packages GLM models with workflow integrations and verticalized tooling for sectors such as finance, healthcare, and manufacturing, grew 248.8 percent year over year to RMB 166 million. A consumer subscription business through Zhipu Qingyan and chat.z.ai contributes a smaller but growing share of revenue.[32]
Z.ai operates inside a fiercely competitive Chinese AI sector and is consistently grouped with five other AI Tigers: Moonshot AI, Baichuan Intelligence, MiniMax, StepFun, and 01.AI.[33] Outside the Tigers, the company competes with the foundation model labs of Chinese internet majors and with stand-alone open-weight labs.
| Competitor | Notable models | Differentiator |
|---|---|---|
| DeepSeek | DeepSeek V3, R1, V4 | Aggressive cost efficiency and reasoning-first releases |
| Alibaba Qwen | Qwen2.5, Qwen3, Qwen3-Max-Thinking | Widest range of sizes from 0.5B to hundreds of billions, multilingual leadership |
| Moonshot AI | Kimi K1.5, K2, K2.5 | Long-context and agent swarms with parallel sub-agents |
| Baichuan Intelligence | Baichuan series | Strong base models, slower agent push |
| MiniMax | MiniMax-Text, MiniMax-Video | Multimodal and video, IPO in 2026 |
| StepFun | Step series | Multimodal and video, Tencent-backed |
| Baidu Ernie | Ernie 4.5, X1 | Vertical integration with search and cloud |
| ByteDance Doubao | Doubao 1.5, 1.6 | Consumer chatbot scale and TikTok integration |
Industry analyses in 2025 and 2026 consistently placed Z.ai's GLM models among the top open-weight model lines globally. After GLM-5 launched in February 2026, Chinese labs collectively held several of the top positions in open-weight AI according to widely cited leaderboards, with GLM-5 (Z.ai), Qwen (Alibaba), Kimi (Moonshot AI), and DeepSeek each leading in different capability dimensions.[15]
The U.S. Entity List designation in January 2025 cut Zhipu off from Nvidia H100 and A100 GPUs and the broader U.S. accelerator supply chain. The company responded with a sustained push toward Huawei Ascend and Cambricon hardware. GLM-4.6's release in September 2025 included the first FP8 and Int4 quantization deployments on Cambricon. GLM-Image in January 2026 became the first state-of-the-art Chinese AI model trained end-to-end on Huawei Ascend processors using the MindSpore framework.[24] GLM-5 followed with a full training run on Ascend hardware. The achievement validated that China's domestic semiconductor and software stack could carry frontier-scale workloads, and made Z.ai a frequent reference in Chinese state media discussions of compute self-reliance.[23]
Z.ai has built out regional headquarters, subsidiaries, and research centers in the United States, the United Kingdom, France, Singapore, and Malaysia. In June 2025 OpenAI said its analysts had seen "notable progress" by Zhipu in securing contracts with governments and state-owned enterprises across Southeast Asia, the Middle East, and Africa, describing the company as offering sovereign large language model infrastructure, often paired with Huawei hardware, in markets including Malaysia, Singapore, the United Arab Emirates, Saudi Arabia, and Kenya.[18] The company has signed frontier AI safety commitments (the Seoul Commitments) alongside OpenAI, Google, Microsoft, Meta, and Anthropic, pledging to conduct red-teaming, publish frontier model transparency reports, and maintain internal security structures for frontier systems.[34] Z.ai has also offered explicit Claude-to-GLM migration paths for users priced out of or restricted from Anthropic's API.[35]
Z.ai has formed strategic partnerships across multiple industries, including:
Z.ai's founding team and researchers have made major contributions to AI research. Key publications include the original GLM paper at ACL 2022, the GLM-130B paper at ICLR 2023, CodeGeeX at KDD 2023, the survey paper "ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools" in 2024 (arXiv:2406.12793), the GLM-4.5 technical report "GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models" in 2025 (arXiv:2508.06471), and the GLM-5 report "GLM-5: from Vibe Coding to Agentic Engineering" in 2026.[5][6] The research output has been organized through the THUDM and zai-org GitHub organizations, which host open-weight releases for GLM, ChatGLM, CodeGeeX, CogVLM, CogView, CogVideoX, GLM-V, and supporting tooling.[3]
CEO Zhang Peng has stated that the company's mission is to achieve "super cognitive intelligence beyond human level" and has frequently described Z.ai as an AGI lab rather than a pure commercial AI vendor.[9] An internal saying at the company holds that "no matter how much money we raise or how much money we make, it will be a hindrance on our road to AGI," framing both fundraising and revenue as means rather than ends. In September 2025 Zhang publicly cautioned that full artificial superintelligence remained unlikely by 2030, though he expected artificial general intelligence to emerge in limited domains by then, a more measured timeline than some Western lab leaders have offered.