China's Z.ai lab has introduced GLM-5.2 — a large-scale open-weights text language model under the MIT license. Thanks to its Mixture of Experts (MoE) architecture and support for a 1 million token context window, the model claims the status of the most powerful open-weights solution for text tasks.
What Happened
Z.ai lab released the GLM-5.2 model, built on an MoE architecture with a total of 753 billion parameters, of which 40 billion are active per request. According to Intelligence Index v4.1 data, the new model has become the leader in the open-weights segment, surpassing projects such as MiniMax-M3 and DeepSeek V4 Pro.
Context
The development of open-weights models is aimed at creating alternatives to closed proprietary systems. The use of Mixture of Experts architecture allows for efficient distribution of computational resources, ensuring high performance at controlled costs, while support for ultra-long context places the model on par with top commercial solutions.
Why It Matters for the Industry
The release of GLM-5.2 intensifies competition with proprietary models like GPT-5.5 and Claude, especially in the field of coding. This forces vendors to rethink pricing and feature availability, and accelerates the blurring of the line between closed SOTA systems and open solutions in high-complexity tasks (reasoning and coding).
Why It Matters for Users
For developers and users, GLM-5.2 provides access to world-class tools for working with massive amounts of data and complex programming. The model can be deployed locally or used via available API gateways (e.g., OpenRouter), significantly lowering the barrier to entry and the cost of creating high-performance AI agents.
What Is Not Yet Known / Limitations
There is a difference in risk assessments: while the technical community focuses on performance, legal experts point to potential compliance issues and risks related to the composition of the training data.
Sources
Author
Look at AI, Editorial Team