Chinese AI giant Zhipu AI releases GLM-5.2: 1 million tokens in context and open-source code
Chinese AI startup Zhipu AI has officially unveiled its new flagship language model — GLM-5.2. This is an open-source solution designed for long-horizon agent tasks and complex programming. The model's key feature is a context window of up to 1 million tokens, opening up fundamentally new possibilities for processing large volumes of data without losing coherence.
The development is distributed under the MIT license, making it as accessible as possible for commercial and research use. The model supports local deployment, and a card with parameters has already been published on the Hugging Face platform: 753 billion parameters, text generation in English and Chinese.
Architectural Innovations and Performance
GLM-5.2 offers a flexible "reasoning intensity" system, allowing users to choose a balance between response quality and latency. The architecture incorporates two key components: the IndexShare mechanism, which reuses one indexer for every four layers of sparse attention, reducing operations per token by 2.9 times, and an updated MTP layer for speculative decoding, increasing the confirmation length to 20%.
According to my data, these improvements directly impact the model's speed and efficiency in real-world scenarios, especially when processing long sequences.
Benchmarks and Leadership
In the FrontierSWE, PostTrainBench, and SWE-Marathon tests, GLM-5.2 confidently outperformed all other open-source models. In standard programming performance benchmarks, it also took first place among open solutions. This confirms that Zhipu AI has managed to create not just a "raw" model, but a full-fledged tool for engineering tasks.
The model is available for local deployment via SGLang, vLLM, Transformers, KTransformers, and Docker Model Runner. Quantizations are supported for llama.cpp, Ollama, and LM Studio, simplifying integration into existing infrastructures.
In my opinion, GLM-5.2 is a significant step forward for open-source AI. The Chinese team is not just catching up with Western developments but is setting new standards in the field of contextual memory and agent systems. If the model confirms its stated characteristics in independent tests, we will see a wave of integrations in the corporate sector in the coming months.