Unsloth's extreme compression of the 753B model using GLM-5.2 enables smooth local deployment and operation on Mac.

According to CoinWorld, Unsloth AI announced that it has compressed the size of Zhipu AI's 753B parameter large model GLM-5.2 by more than 80% using dynamic quantization technology, and released a GGUF format version that supports local deployment on Mac. Through dynamic 1-bit and 2-bit quantization, the original 1.51 TB model can be reduced to 217 GB (1-bit variant) to 239 GB (2-bit variant), allowing ordinary developers and small and medium-sized enterprises to deploy and run it locally offline using only a single Mac Studio. The quantized version achieved a smooth speed of 21.6 tokens/s on a Mac Studio M3 Ultra (256 GB unified memory) device, while retaining 76% to 82% of the original model's accuracy. Currently, the GLM-5.2 GGUF weights are available for download on the Hugging Face platform, and users can load and run them directly through llama.cpp or Unsloth Studio.