Chinese AI giant Zhipu AI releases GLM-5.2: 753 billion parameters and a context of 1 million tokens
Chinese tech startup Zhipu AI has officially unveiled its new flagship language model — GLM-5.2. This open-source solution, distributed under the MIT license, is designed for long-term agent tasks and complex programming. Its key feature is a context window of 1 million tokens, allowing the model to process vast amounts of data without losing coherence.
According to specifications published on Hugging Face, GLM-5.2 has 753 billion parameters and supports text generation in both English and Chinese. The model architecture includes multiple levels of "reasoning intensity," giving users flexibility in choosing between response quality and processing speed (latency).
Technical innovations and performance
At the core of GLM-5.2 are two key innovations: the IndexShare mechanism and an updated MTP layer for speculative decoding. Developers claim that IndexShare allows reusing one indexer for every four layers of sparse attention, reducing operations per token by 2.9 times. Meanwhile, the improved MTP increases confirmation length by up to 20%, which is critical for tasks requiring accuracy and consistency.
In benchmarks measuring real-world programming and agent task-solving capabilities — FrontierSWE, PostTrainBench, and SWE-Marathon — GLM-5.2 confidently outperformed all existing open-source models. It also became the most powerful open model in standard programming performance tests.
Availability and deployment
The model is available for local deployment and supports a wide range of tools: SGLang, vLLM, Transformers, KTransformers, and Docker Model Runner. Additionally, quantizations for llama.cpp, Ollama, and LM Studio are provided for users working with limited computing resources.
My analysis: The release of GLM-5.2 sends a strong signal to the entire AI market. Zhipu AI is not just catching up with Western competitors but, in some aspects — especially in handling long contexts and agent scenarios — is moving ahead. The open MIT license and support for local deployment make this model extremely attractive for developers seeking independence from cloud providers. However, it is worth noting that 753 billion parameters require significant hardware resources, which may limit its widespread adoption in the near term.