Crypto news

Chinese startup Zhipu AI has officially launched its latest language model, GLM-5.2, designed for long-agent tasks and programming. This open-source solution, distributed under the MIT license, offers a context window of 1 million tokens and supports local deployment.

The model, available on the Hugging Face platform, is designed for text generation in English and Chinese and has 753 billion parameters. This makes it one of the largest open-source models on the market.

The GLM-5.2 architecture includes multiple levels of "reasoning intensity," allowing users to choose between response quality and latency. Key innovations include the IndexShare mechanism and an updated MTP (Multi-Token Prediction) layer for speculative decoding. IndexShare 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 up to 20%.

In three key benchmarks — FrontierSWE, PostTrainBench, and SWE-Marathon — GLM-5.2 outperformed all other open-source models. In standard programming performance tests, it also took leading positions among open-source solutions.

For local deployment of GLM-5.2, support is claimed for the SGLang, vLLM, Transformers, KTransformers, and Docker Model Runner frameworks. Quantizations for llama.cpp, Ollama, and LM Studio are also available, making the model flexible for use on various hardware.

My expert opinion: The release of GLM-5.2 is a significant step forward for open-source AI, especially in the programming segment. The 1 million token context opens up new possibilities for code analysis and working with large projects. However, such a large model (753 billion parameters) requires significant computational resources, which may limit its widespread adoption. Nevertheless, the MIT license and support for local deployment make GLM-5.2 an attractive alternative for developers seeking independence from cloud providers.