GLM-5.2 from Zhipu AI: a Chinese open-source model with a context of 1 million tokens and an architecture for agent tasks
Chinese startup Zhipu AI has unveiled its new flagship language model — GLM-5.2. This is an open-source solution that immediately attracted market attention due to its context window of 1 million tokens. The model is designed for long agent scenarios and programming tasks, making it a serious tool for developers and researchers.
GLM-5.2 is distributed under the MIT license and supports local deployment. The model size is 753 billion parameters. It is intended for text generation in English and Chinese, as confirmed by its Hugging Face card.
Architectural Innovations and Performance
One of the key features of GLM-5.2 is support for multiple levels of "reasoning intensity." This allows for flexible balancing between response quality and latency, adapting to specific tasks. The architecture incorporates IndexShare mechanisms and an updated MTP layer for speculative decoding.
IndexShare, according to the developers, 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 20%. These optimizations make the model more efficient when working with large volumes of data.
Benchmarks and Market Positioning
In three key tests — FrontierSWE, PostTrainBench, and SWE-Marathon — GLM-5.2 outperformed other open-source models. In standard programming performance benchmarks, it also took leading positions among open solutions. This confirms its potential for coding automation and complex agent chains.
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 integration into various infrastructures.
My expert commentary: GLM-5.2 is not just another open-source model, but a well-thought-out tool for working with long contexts and agent scenarios. Reducing operations per token by 2.9 times is a significant step forward for local deployment. However, given the size of 753 billion parameters, efficient operation will require powerful hardware. It will be interesting to see how the model performs in real-world projects, especially compared to closed-source alternatives.