Zhipu AI

Zhipu AI (智谱 AI), which markets its products under the name Z.ai, is a Chinese frontier-model company founded in 2019 as a spinout from Tsinghua University's Knowledge Engineering Group (KEG). It develops the GLM (General Language Model) family, which includes ChatGLM, GLM-4, and GLM-4.5, and is one of the cohort of well-funded domestic frontier-model startups sometimes described in Chinese reporting as China's "AI tigers." Most concrete claims about the company rest on scattered secondary reporting rather than primary technical disclosures, in contrast to DeepSeek (arXiv papers) or Qwen (technical report), so coverage here is medium-confidence.

Overview

Zhipu emerged alongside Moonshot AI, Baichuan, and MiniMax in the group of domestic frontier-model startups described in Chinese reporting as China's "AI tigers." Unlike Moonshot (founded 2023) or DeepSeek (hedge-fund-incubated), it is the longest-running of the Tsinghua-adjacent labs, founded in 2019 out of the university's Knowledge Engineering Group. Its GLM (General Language Model) series predates the ChatGPT era and has run in parallel with the Western GPT family. The company is frequently cited in Chinese reporting as competitive with OpenAI on internal benchmarks, though these claims are rarely independently corroborated (Source: epochai.substack.com).

Zhipu operates closer to a closed-API business model than DeepSeek or Alibaba, a positioning consistent with its older, revenue-seeking posture and its enterprise and government customer base. That posture showed in results: Bloomberg reported on July 17, 2026 that Z.ai was set to become the first Chinese AI firm to reach $1 billion in annual sales (Source: bloomberg.com).

Products and models

• GLM-5.3 — launched August 14, 2026 as a post-training-only revision of the GLM-5.2 base model. Z.ai reported a 50% gain over GLM-5.2 on its in-house code benchmark and cyber results it said exceeded its own expectations — CyberGym 84.5% against GLM-5.2's 77.2%, ExploitBench 54.4% against 24.4% — and said it would hold the open weights for roughly two weeks after launch pending safety evaluation and hardening, the first delay of a GLM weight release announced on safety grounds (Source: z.ai; Source: reuters.com).
• GLM-5.2 — open-weights flagship released June 16, 2026 under an unrestricted MIT license, built for long-horizon coding and agentic work: a 753-billion-parameter model with a 1-million-token context window (up to 128K output tokens) and selectable reasoning-effort modes, reported to work out of the box with agentic coding harnesses including Claude Code, Cline, OpenCode, Goose, and Crush. Z.ai reported it scoring 62.1 on SWE-bench Pro (ahead of GPT-5.5) and 74.4% on the FrontierSWE long-horizon coding test (near Claude Opus 4.8), at API pricing of $1.40 per million input tokens and $4.40 per million output tokens, described as roughly six times cheaper than GPT-5.5 and Claude Opus 4.8 (Source: venturebeat.com; Source: z.ai). Z.ai positioned the unrestricted MIT license ("no regional limits") as a contrast with the US export-control directive that prompted Anthropic to take Claude Fable 5 and Mythos 5 offline. Its full open-weights release marks a shift from the partial open-weight posture of the GLM-4 generation. After the June 2026 US restrictions on Anthropic's Mythos-class models, security researchers said GLM-5.2 can match the latest US models at finding software security bugs, though it still trails Anthropic's and OpenAI's systems on other tasks; businesses were reported to be exploiting the narrowing gap to cut costs, with companies including Microsoft weighing how to offer Chinese models (Source: wsj.com). See Export Controls (AI).
• GLM-5.3 — released August 14, 2026 as a post-training-only revision of the GLM-5.2 base model, with weights published on Hugging Face on or before August 28 after a delay Z.ai attributed to cyber-capability evaluation. The weights shipped under a bespoke GLM-5.3 licence rather than MIT — the family's first departure from permissive terms (GLM-5.3 License). The licence is MIT-shaped except for a clause requiring a licensee whose group runs a Model-as-a-Service business and whose group revenue exceeds $10 billion over any consecutive 12 months to pass a Z.AI security review "before using the Software or its derivative works for any commercial purpose," with the review's scope and method "reasonably determined by Z.AI."
• GLM-5 / GLM-5.1 — the GLM-5 line preceding GLM-5.2. GLM-5 (February 2026) scaled the mixture-of-experts backbone to 744B total / 40B active parameters with 28.5T-token pre-training and DeepSeek Sparse Attention, released under the MIT License and aimed at complex systems engineering; GLM-5.1 (April 2026) focused on long-horizon agentic work, reportedly able to run independently for up to eight hours (Source: z.ai; Source: z.ai).
• GLM-4 generation (GLM-4 through GLM-4.7) — the earlier flagship family. GLM-4 (2024) was an API-only dense model with a 128K context and an "All Tools" agent; from GLM-4.5 (July 2025) the flagship moved to a 355B/32B mixture-of-experts design released openly under the MIT License, followed by GLM-4.6 (September 2025, 200K context) and GLM-4.7 (December 2025). Variants include GLM-4-Air (efficient), GLM-4V / GLM-4.5V / GLM-4.6V (vision), and GLM-4-9B (open-weight). GLM-4.5 introduced hybrid thinking / non-thinking modes described as competitive with DeepSeek-R1 and the Qwen3 thinking mode.
• ChatGLM — consumer chatbot and API platform; among the earliest Chinese ChatGPT-class products after Baidu's ERNIE Bot.
• CodeGeeX — code generation model.
• CogView / CogVideoX — image and video generation models.
• BigModel.cn — developer API platform.

The company's open-weight posture is partial: smaller GLM variants are released open, while the flagship is offered API-only. This sits between DeepSeek's fully open-weight flagship and Alibaba/Qwen's Apache 2.0 release at 235B, and contrasts with Moonshot's research-only CC BY-NC-ND license.

In his September 2026 congressional-briefing remarks, Nathan Lambert gave Z.ai's GLM-5.3 (45) and GLM-5.3-Flash (42) as two of the top three Chinese open models on the Artificial Analysis Intelligence Index as of September 14, 2026, and cited anecdotes of AI-industry staff using GLM-5.3 as an alternative to Claude or GPT for speed, lower prices, customization and privacy. He also reported that top Chinese labs including Z.ai and Moonshot AI, which earlier in 2026 preferred building data workflows in-house, had by the summer begun buying challenging reinforcement-learning environments for agentic tasks from established American companies and new Chinese startups (The current balance of power in open models (Nathan Lambert, September 2026)).

ZCode workspace uploads

In a packet-capture analysis published September 18, 2026, the developer Ferstar reported that Z.ai's ZCode desktop coding client silently packaged, encrypted and uploaded entire workspaces, including full .git history, to Aliyun OSS under a public key supplied by the server; in one snapshot manifest the .git directory accounted for 86.6% of the payload. Z.ai open-sourced the client on September 21, 2026 alongside a statement on X, citing the China Academy of Information and Communications Technology as confirming that the zcode-prod bucket holds zero data and NSFOCUS as confirming that the bucket and its objects were deleted. Ferstar's follow-up review of the published repository found two commits covering 6,973 files and 1.03 million lines, with pull requests locked and issues disabled, and found the checkpoint implementation to be purely local Git diffs, contradicting Z.ai's earlier explanation that full-repository uploads were needed for session checkpoint rollback. Ferstar notes that a post-incident audit cannot establish what happened to data uploaded before September 18 (Source: blog.ferstar.org).

Funding

Reported funding in 2024 came from Alibaba, Tencent, Xiaomi, and state-backed funds, accompanying the release of the GLM-4 family and the launch of the flagship API.

Reported first-half 2026 revenue

Z.ai's first-half 2026 revenue of 954 million yuan ($142 million), of which 825 million yuan came from open platform and API sales, became public on 1 September 2026. The report rests on a single subscription briefing whose body could not be retrieved, and the figures are not verified beyond that account (Source: theinformation.com).

People

• Tang Jie (唐杰) — CEO and Tsinghua professor. He is cited for a remark on the US-China gap: "The truth may be that the gap is actually widening" (Source: epochai.substack.com). The quote runs counter to the more common narrative of Chinese labs closing the gap rapidly after DeepSeek-R1, and is treated in reporting as a Chinese-insider data point against rapid convergence.
• Li Peilin — President and co-founder (attribution pending independent confirmation).

Contact details are not reproduced, per privacy conventions.

Safety approach and regulatory environment

Zhipu operates under China's layered generative-AI regime:

• China — Interim Measures for the Management of Generative AI Services (Aug 2023) — requires model registration, content controls, and security assessments for public-facing services. ChatGLM was among the first registered generative-AI services.
• China — Provisions on the Administration of Deep Synthesis Internet Information Services (Jan 2023) — governs synthetic-media outputs, including image and video generators.
• China — Internet Information Service Algorithmic Recommendation Management Provisions (Mar 2022) — covers recommendation surfaces on ChatGLM-powered consumer apps.
• Oversight by the CAC plus six coordinating bodies (MIIT, MPS, MOST, and others).

The company's public posture on AI safety emphasizes alignment with the "socialist core values" required under the Interim Measures rather than the frontier-safety and catastrophic-risk framing common at US and UK labs, a pattern reported across Chinese domestic labs (inference from regulatory context; confidence: medium).

After reports that the Cyberspace Administration of China was investigating potential data-security breaches involving data leaks to Anthropic, Z.ai's Hong Kong-listed shares fell 12.4% on September 23, 2026, against falls of 4% for MiniMax and 4.4% for Alibaba and a 1% dip in the Hang Seng Index (Source: theinformation.com).

Compute infrastructure and custom-chip inquiries

Zhipu completed construction of a 1-gigawatt data center housing only Chinese chips, a milestone reported July 20–21, 2026; the facility has begun partial operations with several computing clusters of more than 10,000 chips each (Source: bloomberg.com; theinformation.com). See AI Data Centers.

The Information reported on July 7, 2026 that Zhipu has made preliminary inquiries with Chinese chip-design houses about a bespoke processor optimized for its GLM models, driven by a 27× surge in token usage for its GLM-5.2 model and by US export controls; Zhipu had not selected a designer (Source: theinformation.com). The inquiries parallel DeepSeek's reported development of its own AI chip the same week.

Export-control status

Zhipu was added to the US Entity List in January 2025 alongside several Chinese AI firms, restricting access to US chips and software, which placed it in the same export-control category as SMIC and other PRC national-champion technology entities (confidence: medium — the specific Entity List designation and date should be independently re-verified). The designation is treated as a concrete test case in export-control and semiconductor supply chain debates over whether such constraints materially affect frontier training, and places Zhipu within US-China chip-access geopolitics discussed under AI race dynamics. Tang Jie's "gap widening" remark is cited in the related US-China AI competition discussion, and GLM-4.5's move to reasoning-mode dual inference is noted alongside Qwen3 and DeepSeek-R1 in the fast-follow discussion.

Z.ai was among the firms — with Alibaba and ByteDance — that Chinese authorities met over the month before July 7, 2026 about restricting overseas access to China's most advanced AI models, including unreleased ones; the Ministry of Commerce-led discussions also covered criminalizing theft or leaks of proprietary AI technology and new rules on funding domestic AI startups (Source: reuters.com). Such curbs would gate overseas access to Z.ai's open-weight GLM line from the Chinese side. See Export Controls (AI).

Comparison to other Chinese labs

History

Relationships

• related: The current balance of power in open models (Nathan Lambert, September 2026) — GLM-5.3 scores and the shift to buying RL-environment data (Lambert, September 2026)
• related: DeepSeek, Alibaba / Qwen Team, Moonshot AI, Baidu (ERNIE), ByteDance / Doubao, China and the US Are Running Different AI Races, AI Race Dynamics, Export Controls (AI)
• instance-of: General-Purpose AI (GPAI) (as a GPAI developer)
• depends-on: China — Interim Measures for the Management of Generative AI Services, China — Provisions on the Administration of Deep Synthesis Internet Information Services, Cyberspace Administration of China (CAC)