What Is Z.ai? Zhipu AI, GLM-5.2 & Open Weights, Explained

Z.ai is the international brand of Zhipu AI, a Chinese lab spun out of Tsinghua University that builds the open-weights GLM model family. It rebranded internationally in July 2025, listed on the Hong Kong Stock Exchange in January 2026, and its June 2026 flagship, GLM-5.2, currently sits at the top of the independent Artificial Analysis index among open-weights models. But here is the part the vendor pages skip: on standard coding tasks, quality has converged. In our own executed benchmark GLM 5.2 scored a perfect 9/9 - the same as nine other models, including GPT-5.5 at about 4.4x the cost. So GLM's real edge is not raw pass rate; it is open weights and long-horizon work.

What Z.ai is

Z.ai is the international brand of Zhipu AI, a Chinese AI lab that builds the open-weights GLM (General Language Model) family. If you have seen the name Z.ai on a model card, a pricing page, or a benchmark leaderboard and wondered whether it is a new company, a wrapper, or a rebrand - it is a rebrand. Same lab, same models, new global name. Its current flagship, GLM-5.2, is an open-weights model that as of July 2026 tops the independent Artificial Analysis index among open-weights systems.

From Zhipu AI to Z.ai

Zhipu AI was founded in 2019 as a spinoff from Tsinghua University's Knowledge Engineering Group, co-founded by Tang Jie and Li Juanzi. In July 2025 the company rebranded internationally to Z.ai and changed its legal name to Knowledge Atlas Technology. On January 8, 2026 it completed an IPO on the Hong Kong Stock Exchange - described as China's first major LLM company to go public - with a market cap reported above HK$52bn (~US$6.6bn). Pre-IPO it raised over US$1.2bn from backers including Alibaba, Tencent, Meituan, Xiaomi and Saudi Arabia's Prosperity7 Ventures. (Treat the financial figures as approximate; several trace to search summaries rather than filings.)

The GLM model family

Z.ai's product is the GLM line, and it has shipped quickly. The short timeline:

Two things stand out. First, the cadence: a new flagship every few months, so any claim you read has a short shelf life - always check the date. Second, the license posture: since 2025 Z.ai has released GLM under the free/open-source MIT License, which is unusually permissive for a frontier-class model. We cover the reasoning-tier jump in our GLM-5 review and the flagship itself in the GLM-5.2 review.

GLM-5.2, the flagship

Per Z.ai's own documentation, GLM-5.2 ships with an approximately 1M-token context window, up to 128K output, multiple thinking modes, function calling and MCP tool integration, and strong long-horizon coding. It is widely reported to be open-weights with core weights under MIT - we state that as reported, since we could not confirm the exact license scope of GLM-5.2 specifically from Z.ai's own docs page in this pass. On the independent Artificial Analysis Intelligence Index it leads open-weights models, and VentureBeat reports it beats GPT-5.5 on several long-horizon coding benchmarks at roughly one-sixth the cost. That long-horizon, agentic strength - not standard-task quality - is where GLM-5.2 earns its keep.

We tested GLM 5.2: 9/9 at $1.99

Most write-ups quote Z.ai's own first-party GLM-5.2 numbers. We ran our own. In the DataLLM Lab coding benchmark (July 2026, 13 models, 9 generate-code-then-run-hidden-tests tasks), GLM 5.2 scored a perfect 9/9 at $1.99 per 1,000 tasks, 12.3s average latency - the sixth-cheapest of the 13. The headline finding is not GLM's score; it is that 10 of 13 models also scored 9/9. On standard coding tasks, quality has converged. What did not converge was cost: the same 9/9 spanned an 88x price range, from Qwen3 Coder Next at $0.10 to GPT-5.5 at $8.83.

The takeaway is a decision rule: on ordinary coding, do not pay for the name. GPT-5.5 and GLM 5.2 delivered the same 9/9, but GPT-5.5 cost about 4.4x more per task. That does not mean all ten 9/9 models are interchangeable everywhere - our tasks are standard, not hard agentic or repo-scale work, and that is exactly where GLM-5.2's long-horizon strength is supposed to separate. It means that if standard-task pass rate is what you are buying, GLM 5.2 is priced sensibly, and the models worth a premium have to prove it on the hard stuff. See how it stacks up against the cheapest strong open coder in our GLM-5 vs DeepSeek comparison.

Pricing & access

Z.ai sells GLM two ways. There is a subscription GLM Coding Plan, reported at roughly $18/mo Lite, $72/mo Pro and $160/mo Max, with discounts of about 10% monthly, 20% quarterly and 30% yearly. And there is pay-per-token API pricing, reported around $1.40 per million input tokens and $4.40 per million output, with a cached-input rate near $0.26 per million. Two cautions: these figures come from secondary pricing/SEO sources rather than Z.ai's live pricing page, so verify current numbers before you commit; and the plan structure suits heavy coding-agent use, while the token API suits variable or bursty workloads. Our GLM Coding Plan breakdown walks through which tier fits which usage pattern.

Open weights & self-hosting

The open-weights posture is GLM's real differentiator, and it cuts both ways. On one hand, MIT-licensed weights mean no per-token lock-in and the option to run on your own hardware. On the other, GLM-5.2 is heavy: full weights are reported around 1.51 TB, and a realistic single-box consumer machine is something like a Mac Studio M3 Ultra with 256 to 512GB of unified memory. That is a serious purchase, and those figures are secondary-source, so treat self-hosting as demanding rather than casual. For most teams the practical access path is a hosted API - which is where a gateway helps: you get GLM-5.2 plus 300+ other models behind one key, and can A/B it against DeepSeek or Qwen without re-plumbing. If self-hosting is the goal, weigh the field first in our best open-source LLM guide.

Who Z.ai is for

Choose GLM-5.2 if you want a frontier-class open-weights model with a permissive license, need long-horizon or agentic coding rather than one-shot snippets, or want a hedge against closed-model lock-in. Look elsewhere if you only need standard coding at the lowest cost - Qwen3 Coder Next and DeepSeek V4-Flash matched GLM's 9/9 for a fraction of the per-task price in our test. And treat every version and price above as a July 2026 snapshot: GLM-5 shipped in February and GLM-5.2 in June, so a newer GLM could land at any time. The safest way to keep up is to keep GLM-5.2 one model name away and re-benchmark when the next one drops.

FAQ