Chinese AI company Zhipu AI released GLM 5.2 through its Z.AI platform on June 13, an open-weight model with a 1-million-token context window that the company says outperforms GPT-5.5 and Claude Opus 4.7 on several long-horizon benchmarks.

The model runs on a 744-billion-parameter Mixture-of-Experts architecture. Its 1-million-token context is roughly five times larger than GLM 5.1’s 200,000-token window, which matters for tasks requiring a large codebase, legal document archive or research corpus in a single inference session. It supports up to 131,072 output tokens per response — enough for detailed code generation and structured document creation without truncation.

GLM 5.2 includes a dual thinking-effort system called High and Max, letting developers control how much reasoning the model applies per query. The company reports a score of 81.0 on Terminal-Bench 2.1, compared to 62.0 for GLM 5.1, a benchmark designed to test long-horizon software engineering tasks performed autonomously.

The model’s weights are licensed under MIT, meaning developers can use, modify and redistribute them without restriction. At launch, access was limited to Coding Plan subscribers. The open weights were promised the following week but had not appeared on Hugging Face by launch day, which drew criticism from developers expecting immediate availability.

GLM 5.2 enters a market where every major AI lab is competing on context length and agentic performance. Google’s Gemini 2.5 Flash launched this week claiming four times faster inference than its predecessor. Microsoft’s 10 billion dollar Japan AI investment signals where enterprise compute spending is heading. For developers in Asia, an open-weight model from a Chinese lab with genuine frontier-tier context length offers a real alternative to US-dominated top-tier offerings.

Zhipu’s benchmarks at launch were self-reported without independent third-party evaluation data. Real-world assessments from the developer community will arrive over the coming weeks. Full technical documentation and API access are available at the Z.AI official platform. The open-weight release is also expected on Android XR developer pipelines once the Hugging Face weights go live.

GLM 5.2 is the most capable open-weight model Zhipu has released. The MIT license and 1-million-token context make it worth serious attention from anyone building long-context applications — once independent testing confirms whether the benchmark claims hold up.