GLM-5.2 from Zhipu AI: 1 million token context and open-source code for agent tasks
Chinese startup Zhipu AI has released its new flagship language model GLM-5.2, targeting long agent scenarios and programming tasks. The model is open-sourced under the MIT license and supports local deployment, making it accessible to a wide range of developers.
A key feature of GLM-5.2 is its context window of 1 million tokens. This allows the model to process and analyze vast amounts of data, including entire codebases or long dialogues, without losing coherence. On Hugging Face, the model is listed as generative for English and Chinese, with an impressive size of 753 billion parameters.
The architecture of GLM-5.2 includes multiple levels of "reasoning intensity," giving users flexibility in choosing between response quality and latency. Built-in IndexShare mechanisms and an updated MTP (Multi-Token Prediction) layer enable speculative decoding. IndexShare reuses one indexer for every four layers of sparse attention, reducing operations per token by 2.9 times. The MTP update, in turn, increases confirmation length by up to 20%, accelerating generation.
In three key benchmarks — FrontierSWE, PostTrainBench, and SWE-Marathon — GLM-5.2 confidently outperformed all other open-source models. In standard programming performance tests, it also took the leading position among open-source solutions.
For local deployment, GLM-5.2 supports SGLang, vLLM, Transformers, KTransformers, and Docker Model Runner. Additionally, quantizations are available for llama.cpp, Ollama, and LM Studio, simplifying model integration into various infrastructures.
Analyst opinion: The release of GLM-5.2 is a significant step forward in the open-source AI segment. The combination of a 1 million token context, 753 billion parameters, and the open MIT license creates a powerful tool for developers, especially in automation and programming. However, it is worth remembering that such giant models require significant computational resources, and their real effectiveness will depend on the quality and relevance of the training data. In the long term, this could spur competition among open-source models, benefiting the entire community.