Unsloth’s extreme compression lets a 753B model run on a Mac

Unsloth AI has found a way to stuff a 753-billion-parameter monster onto a Mac Studio. Its new dynamic quantization technique shrinks Chinese AI lab Zhipu AI’s GLM-5.2 model by over 80% — from a jaw-dropping 1.51 TB down to between 217 GB and 239 GB, depending on the quantization level (1-bit or 2-bit UD-IQ2_M). That means developers and small businesses can now run the thing entirely offline on a single machine.

On a Mac Studio M3 Ultra with 256 GB of unified memory, the quantized model chugs along at 21.6 tokens per second while retaining 76% to 82% of the original precision. In Unsloth’s own benchmarks, the locally running 1-bit GLM-5.2 GGUF produced code for a pixel-art Flappy Bird clone (complete with sound effects and particle systems) that was on par with outputs from Claude 4.8 Opus and GPT-5.5.

GLM-5.2 is a Mixture-of-Experts model from Zhipu AI with a staggering 753B total parameters and a 1-million-token context window. Usually, running something that big requires an expensive cloud cluster; Unsloth’s quantization flips the script by drastically lowering the barrier for individuals and small teams. The GGUF weights are already up on Hugging Face, and you can load them via llama.cpp or Unsloth Studio.