Zhipu AI

Updated June 2026

Zhipu AI (internationally branded Z.ai) is the Beijing lab behind the open-weight GLM model family and the Z.ai assistant. Spun out of Tsinghua University in 2019, it is one of China's 'AI tigers', became the world's first listed large-language-model company on its January 2026 Hong Kong IPO, and pairs MIT-licensed open weights with an agentic-coding focus — all while on the US Entity List.

Website →API →Pricing →

Zhipu AI — branded Z.ai internationally — is the Beijing lab behind the open-weight GLM model family and the Z.ai assistant. Spun out of Tsinghua University in 2019, it is one of China’s “AI tigers” and built its reputation on releasing capable models as MIT-licensed open weights while pushing hard into agentic coding. In January 2026 it became, in its own words, the world’s first listed large-language-model company, debuting on the Hong Kong Stock Exchange (Caixin).

Zhipu’s 2026 story is one of contrasts: a Tsinghua research spin-out turned public company; an open-weight champion that competes at the frontier of coding benchmarks; and a US Entity List designee that trains its models on Chinese silicon rather than Nvidia. This guide covers Zhipu as an AI provider — the GLM models, the listing and the money, the strategy, and the risks.

Quick facts

History and founding

Zhipu AI was founded in 2019 as a commercial spin-out of Tsinghua University’s Knowledge Engineering Group (KEG) — the research lab also known as THUDM — by professors Tang Jie and Li Juanzi (Quartz). That academic heritage runs through everything Zhipu ships: its GLM (“General Language Model”) architecture, the open ChatGLM dialogue models that made it well known in 2023, and a continuing flow of research on agents, vision and video. The company is led by CEO Zhang Peng, with Tang Jie as its defining scientific figure.

By 2024 Zhipu was reportedly the largest of China’s AI startups by headcount (around 800 people) and a fixture among the country’s “AI tigers” — the cohort of well-funded large-model startups that also includes Moonshot (Kimi), MiniMax, Baichuan, StepFun and 01.AI (MIT Technology Review).

The IPO and financials

Zhipu’s defining 2026 corporate event was its Hong Kong IPO on 8 January 2026 (stock code 2513.HK), which it billed as the world’s first listing by a large-language-model company. The offer priced at HK$116.20 a share, raised about HK$4.17 billion, was retail-oversubscribed by more than 1,000 times, and valued the company at roughly US$7 billion; around 70% of proceeds were earmarked for model R&D through 2028 (Caixin, PR Newswire).

Before listing, Zhipu had raised roughly US$1.5 billion since founding from an unusually broad investor base — Alibaba, Tencent, Ant Group, Meituan, Xiaomi, Hillhouse, Qiming Venture Partners, Chinese local-government funds, and Saudi Aramco’s Prosperity7 Ventures (Bloomberg). Its debut annual results reported 2025 revenue of about RMB 724 million (~$105M), up roughly 132% year on year, against a widened net loss of about RMB 4.7 billion — an enterprise-heavy mix (on-premises deployments to state-owned enterprises and financial institutions) with steep losses typical of the sector (mlq.ai). As a newly public company its market value has been volatile, swinging sharply around model launches and US policy news.

The GLM model family

Zhipu’s research output centres on the GLM line, most of it released as open weights under the MIT licence for free commercial use and self-hosting:

The current lineup is built around the GLM-5 series of large mixture-of-experts models tuned for agentic engineering — Zhipu’s slogan for the generation is “from vibe coding to agentic engineering.” The open weights live on Hugging Face under the zai-org organisation, alongside the company’s older THUDM research repos. Full benchmarks and pricing are on the GLM model page.

Strategy: open weights, agentic coding, sovereign compute

Zhipu’s strategy rests on three pillars. First, open weights: like DeepSeek and Alibaba’s Qwen, it releases its flagships under permissive licences (MIT), betting on developer adoption and enterprise self-hosting rather than withholding its best models. Second, agentic coding: the GLM-5 series is explicitly optimised for long-horizon software-engineering agents — Zhipu markets GLM-5.1 as sustaining autonomous work for hours across hundreds of iterations — and it sells a developer-friendly GLM Coding Plan that plugs into tools like Claude Code and Cline (Hugging Face).

Third, and most distinctively, sovereign compute: GLM-5 was reportedly trained and served on Chinese accelerators — Huawei Ascend, Moore Threads, Cambricon and Kunlunxin — rather than Nvidia, a direct response to US export controls and a selling point for Chinese government and enterprise buyers (Reuters via AOL).

US Entity List

In January 2025, the US Commerce Department added Beijing Zhipu Huazhang Technology and subsidiaries to its Entity List, citing concerns that the company was helping advance China’s military modernisation through AI; the listing restricts Zhipu’s access to US technology without a (generally denied) licence (SCMP). Zhipu “strongly opposed” the move, saying it “lacks a factual basis” and would not substantially affect its business because it does not rely on US large-model technology (Reuters/US News). The designation is part of why Zhipu leaned into domestic chips and open weights.

Products and ecosystem

• Z.ai assistant — the consumer and developer chat product (formerly ChatGLM), at chat.z.ai.
• BigModel / Z.ai API platform — the model-as-a-service developer platform and console (docs.z.ai).
• GLM Coding Plan — subscription tiers for agentic coding, integrating with Claude Code, Cline, Roo Code and others.
• AutoGLM — an agent that operates phones and apps via natural-language commands.
• Specialist models — CogVideoX (video), CogView/GLM-4V (image and vision) and CodeGeeX (code).

Competition and market position

Zhipu competes in the Chinese open-weight cluster with DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi and MiniMax. Its differentiation is the combination of a strong agentic-coding focus, MIT licensing, Tsinghua research depth, and a public-company balance sheet. On independent leaderboards its models are competitive but not the open-weight leader: Artificial Analysis places GLM-5.1 around the middle of the open-weight pack, behind Kimi K2.6 and DeepSeek V4-Pro on its composite index (Artificial Analysis), even as Zhipu’s own coding benchmarks for GLM-5.2 rank at or near the top of the open field.

Against the US frontier labs — OpenAI, Anthropic and Google — Zhipu trades the last increment of general capability for openness, low cost and sovereign deployment, the same trade its Chinese peers make.

Controversies and risks

• US Entity List. The January 2025 designation restricts US technology access and frames Zhipu within the US-China tech rivalry.
• National Intelligence Law and data residency. Like all Chinese providers, Zhipu is subject to China’s 2017 National Intelligence Law; hosted-API use routes data through Chinese infrastructure. Self-hosting the MIT open weights mitigates this.
• Profitability. Heavy losses against modest revenue — a common critique of China’s AI-tiger cohort, sharpened by Zhipu’s public-market scrutiny.
• Benchmark transparency. Zhipu’s headline “beats GPT-5.5 / Opus” coding claims are vendor-reported; independent indices rank its models lower, so vendor and independent figures should be read separately.
• Content controls. As a China-based provider, hosted GLM aligns with PRC content rules, a consideration for politically sensitive or neutral use.

Recent developments (2026)

• Hong Kong IPO (8 January 2026) made Zhipu the world’s first listed LLM company (2513.HK).
• GLM-5 (February 2026) launched as the open-weight flagship, reportedly trained without Nvidia chips.
• GLM-5.1 (April 2026) sharpened the long-horizon agentic-coding focus.
• GLM-5.2 (13 June 2026) extended the context to 1M tokens and topped several open-weight coding benchmarks on Zhipu’s testing.
• Debut earnings reported 2025 revenue up ~132%, alongside a reported plan for a second listing on Shanghai’s STAR Market.

Where Zhipu excels

• Open weights. MIT-licensed flagships make self-hosting, fine-tuning and air-gapped deployment genuinely viable.
• Agentic coding. The GLM-5 series is purpose-built for long-horizon software-engineering agents, with a developer-friendly coding subscription.
• Sovereign compute. Training on Chinese accelerators is a differentiator for buyers wary of US-controlled hardware and a hedge against export controls.
• Research depth and capital. Tsinghua heritage plus a public listing give it both talent and a funded balance sheet.

Where Zhipu falls short

• Not the open-weight leader. On independent indices its models trail DeepSeek V4 and Kimi K2.6 on the hardest tasks.
• Entity List and geopolitics. US restrictions and China-provider trust concerns limit Western adoption.
• Losses. Large cash burn against modest revenue invites scepticism about the business model.
• Benchmark credibility. Aggressive vendor claims need independent corroboration.

Developer resources

Zhipu serves GLM through the Z.ai / BigModel platform (docs.z.ai), an OpenAI-SDK-compatible API with per-token pricing, tool calling, MCP support, JSON output and context caching. Open-weight GLM models are published on Hugging Face under MIT and can be run with vLLM, SGLang, Transformers, Ollama and LM Studio (including FP8 and GGUF quantisations). The GLM Coding Plan brings the models into agentic IDE tools, and consumer access is via the free Z.ai assistant. Model details and benchmarks live on the GLM model page.

Frequently asked questions

What is Zhipu AI?

Zhipu AI, branded Z.ai internationally, is a Beijing-based artificial-intelligence company spun out of Tsinghua University in 2019. It builds the open-weight GLM model family and the Z.ai assistant, and is one of China’s leading large-model startups.

Is Zhipu the same as Z.ai?

Yes. Z.ai is the international brand of Zhipu AI (Chinese: 智谱). The company also runs the BigModel developer platform and is listed on the Hong Kong Stock Exchange as 2513.HK.

Are Zhipu’s GLM models open source?

Most are released as open weights under the MIT licence, including the GLM-5 series flagships — free to download, self-host, fine-tune and use commercially. Zhipu monetises through its hosted API, the GLM Coding Plan and enterprise deployments rather than by withholding its best models.

Why is Zhipu on the US Entity List?

The US Commerce Department added Zhipu to its Entity List in January 2025, citing concerns it was helping advance China’s military modernisation through AI. The listing restricts Zhipu’s access to US technology. Zhipu disputes the rationale and says it does not rely on US large-model technology.

Is Zhipu safe to use for sensitive data?

Use caution with the hosted Z.ai app and API, which route data through Chinese infrastructure and are subject to China’s National Intelligence Law and content rules. For sensitive or regulated work, the safer route is to self-host Zhipu’s MIT open weights, or use a Western-hosted provider such as Anthropic or Mistral.

How does Zhipu compare to DeepSeek?

Both are Chinese open-weight leaders. DeepSeek is known for the lowest prices and a strong general model under MIT; Zhipu’s GLM line is pitched harder at agentic coding and long-horizon agents. On independent indices DeepSeek V4 currently rates above GLM on the hardest tasks, while Zhipu competes on coding focus and its developer tooling.

Models