The Chinese startup Zhipu AI has released the GLM-5.2, an open source language flagship focused on long agent tasks and programming. The model has 753 billion parameters, supports English and Chinese, and offers a contextual window of 1 million tokens. GLM-5.2 is distributed under the MIT license and supports local deployment through various platforms, including Ollama, vLLM, and llama.cpp.

The model architecture includes a "reasoning intensity" selection function to balance quality and speed, as well as IndexShare and MTP optimization. IndexShare reduces the number of transactions per token by 2.9 times, and the updated MTP layer increases the confirmation length to 20%. In key development benchmarks such as FrontierSWE, PostTrainBench and SWE-Marathon, as well as standard programming tests, GLM-5.2 outperformed other open-source solutions. The release took place against a background of scandal surrounding another open model Rio 3.5 Open, which the Nex team had identified as a direct merger between Nex-N2-Pro and Qwen3.5-397B-A17B without full training.