Chinese startup Zhipu AI has introduced GLM-5.2: a context of 1 million tokens and an open MIT license.
Chinese company Zhipu AI has released its flagship language model GLM-5.2, designed for long agent scenarios and programming tasks. The open-source solution features a context window of 1 million tokens, an MIT license, and support for local deployment.
Technical Specifications and Architecture
According to the specification on Hugging Face, GLM-5.2 is designed for text generation in English and Chinese. The model has 753 billion parameters. Developers have implemented several levels of "reasoning intensity," allowing flexible balancing between response quality and latency. The architecture also integrates IndexShare mechanisms and an updated MTP layer for speculative decoding.
IndexShare reuses a single indexer for every four layers of sparse attention, reducing operations per token by 2.9 times. The MTP update increases confirmation length by up to 20%, significantly accelerating inference in complex scenarios.
Performance and Benchmarks
In three key tests — FrontierSWE, PostTrainBench, and SWE-Marathon — GLM-5.2 outperformed all existing open-source models. In standard programming performance benchmarks, the Chinese development also took the leading position among open-source counterparts.
Availability and Deployment
GLM-5.2 is distributed under the open MIT license. For local deployment, support is claimed for SGLang, vLLM, Transformers, KTransformers, and Docker Model Runner. Additionally, quantizations are available for llama.cpp, Ollama, and LM Studio, making the model suitable for use on consumer hardware.
Expert opinion: The release of GLM-5.2 marks an important milestone in the open-source AI model race. The 1 million token context and MIT license are a direct challenge to proprietary solutions from OpenAI and Google. However, the key question remains the practical applicability of such a model on standard hardware: 753 billion parameters require significant computational resources, even with quantization. The market awaits independent performance tests in real-world agent programming tasks.