Leading domestic large model vendor Zhipu AI has recently officially released and open-sourced its new model GLM-4.6. According to the official introduction, the model has achieved significant improvements in key capabilities such as Agentic Coding. Its programming capability has aligned with international top models Claude Sonnet4 on public benchmarks and real programming tasks, and it has surpassed DeepSeek-V3.2-Exp, becoming the strongest code generation model in China so far.

This release not only marks a technological breakthrough but also achieves a milestone in domestication adaptation. Zhipu announced that GLM-4.6 has successfully implemented FP8+Int4 mixed quantization deployment on Cambricon domestic chips, which is the first production-ready chip-integrated solution of this mode.

In addition, the model can run stably at native FP8 precision on Molybdenum Thread’s new generation GPU based on the vLLM framework. This marks that domestic GPUs now have the ability to collaborate and iterate with cutting-edge large models, laying a solid foundation for building an independent and controllable AI ecosystem.

The release of GLM-4.6 not only strengthens Zhipu's leading position in the field of code large models, but also provides Chinese AI developers with more powerful and efficient tools, while promoting the deep integration and development of AI large models with domestic hardware.