Overview
GLM-5.2 is Zhipu's strongest model to date, released on 2026/6/13 and **officially open-sourced under the MIT license on 2026/6/17**, representing the next generation of "mythical open-source." **It has been tested to handle a complete engineering file of 880,000 tokens at once**, ranking **first among open-source models in the Code Arena global frontend code evaluation**, with overall strength second only to the closed-source Claude Fable 5 (which is subject to US export controls).
Technically, it relies on the **self-developed IndexShare architecture + optimized MTP speculative decoding module**, with a new "thinking gear" for adjustable reasoning depth. **Terminal-Bench programming tests significantly surpass the previous generation**; **FrontierSWE long-horizon task benchmarks place it between Claude Opus 4.7 and 4.8**—this is the first time a domestic open-source model has entered the international top-tier closed-source flagship range in long-horizon programming benchmarks.
The open-source license is MIT, and the weights are available on Hugging Face / ModelScope. **On the first day, it completed adaptation for three domestic chips: Huawei Ascend, Moore Threads, and T-Head**—pure domestic computing power for private deployment is ready out of the box. The accompanying **ZCode 3.0 programming tool** simultaneously switches to the self-developed ZCode Agent kernel, end-to-end unleashing GLM-5.2's long-horizon capabilities.
Zhipu explicitly stated regarding the Claude Fable 5 ban due to US export controls within 4 days: "**Frontier intelligence should not belong only to a few, nor should it be revoked at any time by a few rules. It should be open, available, buildable, and serve every developer**"—turning MIT open-source + 1M context + the strongest domestic Coding directly into a counter-proposition to regulatory supply.
Key Features
- Officially Open-Sourced Under MIT License: Open-sourced simultaneously on Hugging Face / ModelScope on 2026/6/17, commercially usable, fine-tunable, self-hostable, the most generous license among major domestic companies
- Tested Engineering Capacity of 880,000 Tokens: Loads a complete engineering file of 880,000 tokens at once while maintaining effective retrieval and reasoning, with 1M context available rather than just a paper parameter
- First in Code Arena Among Open-Source Models: Achieves first place among open-source models in the global frontend code evaluation Code Arena, overall second only to closed-source Fable 5
- FrontierSWE Long-Horizon Benchmark Ranks Among International Top Tier: FrontierSWE long-horizon tasks place it between Claude Opus 4.7 and 4.8, the first time a domestic open-source model has entered the international top-tier closed-source flagship range
- IndexShare + MTP Speculative Decoding: Self-developed architecture reduces long-context computation, combined with MTP speculative decoding to significantly improve inference throughput
- Adjustable Thinking Gear: New thinking gear function allows switching reasoning depth based on task difficulty, similar to reasoning_effort but more granular
- First-Day Adaptation for Three Domestic Chips: Completed adaptation for Huawei Ascend, Moore Threads, and T-Head on the first day, enabling out-of-the-box private deployment with pure domestic computing power
- ZCode 3.0 End-to-End Integration: Accompanied by ZCode 3.0 self-developed Agent kernel, with deep optimization for long-horizon reasoning, tool calling, and large-scale engineering pipelines
Use Cases
- Long-horizon programming tasks (dependency updates, cross-file refactoring, long program generation)
- Whole-repository code understanding, long paper analysis, ultra-long conversations, and other 1M context scenarios
- Enterprise AI platform selection with requirements for domestic, compliant, or privatized deployment
- Developers looking to avoid regulatory uncertainty of overseas models (following the Fable 5 incident)
- Private deployment on pure domestic computing power (Huawei Ascend/Moore Threads/T-Head)
- Academic research: model fine-tuning and alignment studies under the MIT license
Pros
- Triple package: MIT open-source license + 880,000-token engineering test + first in Code Arena among open-source models
- FrontierSWE long-horizon benchmark first brings domestic open-source into the international top-tier closed-source range
- First-day adaptation for three domestic chips, enabling private deployment on pure domestic computing power
- End-to-end integration with ZCode 3.0 self-developed Agent kernel
- No-cost upgrade for GLM Coding Plan subscribers
- In the context of Fable 5 being taken down due to regulation, 'open and available' is a hard-to-replace product positioning
Pricing
**Open-source version**: Open-sourced on HuggingFace/ModelScope under the MIT license on 6/17, commercially usable, fine-tunable, self-hostable. **GLM Coding Plan** (subscription): Four tiers: Lite / Pro / Max / Team, with full push of GLM-5.2 starting 6/13, no additional cost for subscribers. **API**: Available via chatglm.cn, specific pricing subject to Zhipu's official website. **Domestic chip deployment**: First-day adaptation for Huawei Ascend, Moore Threads, and T-Head, deployable on a pure domestic computing power stack.
Summary
GLM-5.2 is the most significant event for domestic open-source large models in June 2026—**MIT open-source + 880,000-token engineering test + first in Code Arena among open-source models + FrontierSWE entering the international top-tier closed-source range + first-day adaptation for three domestic chips**. Against the backdrop of Claude Fable 5 being banned within 4 days due to US export controls, GLM-5.2 becomes the most convincing counter-proposition of 'open, available, and not revocable.' If you work on long-horizon programming or long-document scenarios, we recommend immediately adopting GLM Coding Plan + ZCode 3.0; if you consider private deployment, you can download weights from HuggingFace starting 6/17; if you are building a compliance-sensitive AI platform for state-owned enterprises, the three domestic chip adaptation is a rare 'truly domestic' option currently available.
Version History
- GLM-5.2 Officially Open-Sourced (MIT) (2026-06-17): MIT license available on Hugging Face / ModelScope; first in Code Arena among open-source models; FrontierSWE between Opus 4.7 and 4.8; first-day adaptation for three domestic chips: Ascend, Moore Threads, T-Head
- GLM-5.2 + ZCode 3.0 Full Push (2026-06-13): Strongest domestic open-source Coding + 1M context, full push to GLM Coding Plan; ZCode 3.0 switches to self-developed Agent kernel
- GLM-5 (First half of 2026): Zhipu's 1.3T MoE flagship, strong multimodal visual encoding capabilities