GLM-5.2 from Zhipu AI: 1 million tokens of context and a revolution in open-source
Chinese startup Zhipu AI has unveiled its new flagship language model GLM-5.2, designed for long-term agent tasks and programming. The open-source solution features a context window of 1 million tokens, an MIT license, and the ability for local deployment — a significant step forward for the industry.
According to data on Hugging Face, GLM-5.2 is a text generation model for English and Chinese with a massive 753 billion parameters. This architecture enables it to handle complex scenarios requiring deep contextual understanding, which is critical for code development and analysis.
The key innovation is support for multiple levels of "reasoning intensity," giving users flexibility in choosing between response quality and latency. The architecture integrates IndexShare and an updated MTP layer for speculative decoding. Developers claim IndexShare reuses one 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 generation.
In the FrontierSWE, PostTrainBench, and SWE-Marathon benchmarks, GLM-5.2 outperformed all other open-source models. In standard programming performance tests, it also took the leading position among open-source solutions. This confirms that Zhipu AI has created not just another model, but a tool that truly surpasses competitors in practical tasks.
GLM-5.2 is distributed under the MIT license, opening up broad opportunities for commercial use. For local deployment, support is announced for SGLang, vLLM, Transformers, KTransformers, and Docker Model Runner. Quantizations are available for llama.cpp, Ollama, and LM Studio — simplifying integration into existing infrastructures.
Analytical commentary: This release marks a shift in the balance of power in the open-source AI market. Chinese developers are increasingly challenging Western giants, offering solutions with impressive technical specifications. However, it is worth noting that 753 billion parameters pose a challenge for local deployment: not every user will be able to run such a model on their hardware. Nevertheless, support for quantization and flexible reasoning modes makes GLM-5.2 attractive for those willing to invest in serious computing resources.