GLM-5.2 Released: Zhipu Opens Up Its Most Powerful Open Source Model – 1M Context, MIT License and Open Source Next Week
Main chat
A chat for vibe coders: news, guides, live cases, marketplace, and finding executors.
On June 13, 2026, at 17:21 Beijing time, Zhipu AI (the international brand Z.ai) opened access to GLM-5.2 for all GLM Coding Plan users - Lite, Pro, Max and Team tariffs. According to the company, this is the most powerful open model in their history. The API will be released next week, at which time the model will be published in open-source under the MIT license.
The release took place exactly seven days after the appearance of the first signals of internal testing in the network - Z.ai adheres to this rhythm with metronome accuracy: GLM-4.7, GLM-5.0, GLM-5.1 all went public 7 days after the start of beta testing.
What is known about the GLM-5? 2
The official information at the time of publication is few words, but specific.
1M context is actually working. Zhipu characterizes GLM-5.2 as a model with "о人о 1Mо" - "a real 1M context." This is an important caveat: many models claim a large contextual window, but in practice quality plummets over long distances. Zhipu is clearly positioning 1M as a feature, not a marketing metric.
Long-Target Leadership. The company says GLM-5.2 “continues to lead long-range tasks.” This is a direct continuation of the GLM-5.1 vector, which focused on long agent scenarios.
MIT license, open source next week. The model will go open source under MIT next week - with no restrictions on commercial use, financial tuning and self-hosting.
API next week. Developers not using the GLM Coding Plan will be able to connect via APIs from next week.
How Zhipu got to GLM-5.2 in 4 months
To understand the significance of this release, you need to understand where it grew.
GLM-5: February Breakthrough
On February 11, 2026, a few days before the Chinese New Year, Z.ai released the GLM-5 — and it was a real break with the previous generation. The model grew from 355 billion parameters (32B active) to 744 billion (40B active). Pre-training dataset – from 23T to 28.5T tokens. The architecture switched to DeepSeek Sparse Attention for efficient long-context work, supporting a 200K token window.
Strategically important fact: the GLM-5 is trained entirely on Huawei Ascend chips using the MindSpore framework – without a single NVIDIA chip. In the face of U.S. export sanctions, the GPU is not just a technical solution, but a political statement.
Zhipu’s shares on the Hong Kong Stock Exchange rose 60 percent in the three days following the release. Simultaneously with the release of the model, the company raised prices for the GLM Coding Plan by 30%.
GLM-5.1: April acceleration
In March, GLM-5.1 was opened to subscribers, and on April 8, it was published in open-source. The main result: a 30% increase in coding quality compared to GLM-5, close to Claude Opus 4.6 on the benchmarks SWE-bench Verified, HLE and BrowseComp.
The withdrawal rate is 68 tokens per second. The cost is $3 per million tokens.
GLM-5.1 High-Speed: May Surprise
In May, Zhipu released a “high-speed” version of GLM-5.1 — 400 tokens per second. Almost 6 times faster than the base. For long-chain agent scenarios, this is a critical characteristic.
GLM-5-Turbo: Specialization in Agents
In March, the GLM-5-Turbo was released in parallel, a model designed from scratch for OpenClaw scripts (complex automated agent workflows). Cost: $1.2 per million input tokens, $4 per million weekend. Five times cheaper than the Claude Opus 4.6.
Why exactly coding
The choice of Z.ai in favor of AI Coding as the main direction is not accidental – it is the result of readable market logic.
Anthropic has reached $30 billion in ARR, largely by developers and Claude Code. The global AI Coding market is at least $100 billion. The Chinese market is just entering an acceleration phase.
Zhipu positions the transition from “Vibe Coding” to “Agentic Engineering” – the completion of complex system engineering tasks and long agent scenarios – as the main thesis of the entire GLM-5 series.
This is indicated by business dynamics: cloud revenue in 2025 increased by 293% year-on-year. The business model is shifting from project contracts to token billing – the same path that Anthropic has taken.
Philosophy of Openness: Not Just Business
The official statement on GLM-5.2 contains a phrase worth quoting in its entirety. Zhipu wrote:
“At a time when advanced models suddenly become unavailable, Zhipu chooses a different path: advanced intelligence should not only belong to a few and should not be withdrawn at any time by a few.” It should be open, accessible, buildable — and serve every developer.”.
This is an indirect but readable reference to Anthropic’s Fable 5 situation, which is restricted to users outside the United States. In the race for developers’ minds, Z.ai is betting on openness as a differentiator.
GLM-5.2 in the context of the Chinese AI market
Zhipu is not the only player on the field. According to IDC, the company ranks third in the Chinese LLM market after Alibaba (Qwen) and Baidu (ERNIE). But in terms of influence in the developer community, open-source strategy actually outperforms market share.
On benchmarks, the GLM-5 (744B / 40B active) is comparable to the Qwen 3.5 (397B / 17B active) and the DeepSeek V4-Pro (1.6T / 49B active). GLM leads in Chinese, DeepSeek in coding, and Qwen in multimodality. Three systems are competitive on the open-weight frontier.
GLM-5.2 with claimed real 1M context and improved long tasks potentially changes this ratio – especially for agent applications, where the length of the context directly determines the complexity of the tasks being solved.
What it means for developers
Right now: If you have a GLM Coding Plan (Lite, Pro, Max or Team), the GLM-5.2 is already available. You can test it.
** Next week, the API will open to all developers. The model will be released on HuggingFace under the MIT license - you can download weights, fine tune, run locally.
1M context is the ability to transfer the entire code base of an average project in one session. Improved Long Tasks: More reliable agent scripts with Claude Code or analogues. Open source – the ability to run a model on your own hardware without depending on the API.
Price: Based on GLM-5 ($1.00/3.20 per million OpenRouter tokens) and GLM-5.1 ($3 per million tokens) – GLM-5.2 is expected in the same price range. 5-8 times cheaper than the Claude Opus level with comparable capabilities.
What's unknown
To be honest, the official announcement of GLM-5.2 is minimalistic. Specific benchmarks — SWE-bench Verified, AIME 2026, GPQA-Diamond — were not disclosed at the time of publication. No architectural changes to GLM-5.1 have been announced. The exact release date of the API and open source is “next week” without a specific number.
Details will appear with the release of the API and the publication of the scales. Independent community tests, as always, will give a more honest picture than internal benchmarks.
Chronology of the GLM-5 series
| Версия | Дата | Ключевое |
|---|---|---|
| GLM-5 | 11 февраля 2026 | 744B параметров, 200K контекст, Huawei Ascend |
| GLM-5-Turbo | Март 2026 | Специализация на агентных задачах |
| GLM-5.1 | Апрель 2026 | +30% к кодированию, 68 tok/s, $3/M токенов |
| GLM-5.1 High-Speed | Май 2026 | 400 tok/s |
| GLM-5.2 | 13 июня 2026 | 1M контекст, лидерство в длинных задачах |
Five major updates in four months. Faster than most Western labs produce one generation.
*The article will be updated after the publication of the API and open-source scales. Follow the official Z.ai and HuggingFace channels. Current June 13, 2026. *
