GLM 5.2 runs on MacBook Pro M5 at 2-2.8 tokens/s

A user reports successfully running the GLM 5.2 model on a MacBook Pro M5 with 48 GB of RAM using the pipenetwork/GLM-5.2-MLX-mixed-3_6bit checkpoint. The inference speed reaches between 2 and 2.8 tokens per second, depending on system load.

• The user utilized Flash MOE to integrate GLM 5.2 into a custom Claude Desktop-style application.
• Benchmarking with a 4.4k pre-prompt resulted in 501 output tokens at 2t/s over 6 minutes and 38 seconds.
• A test without a pre-prompt processed 400 input tokens to produce 337 output tokens in 3 minutes and 17 seconds.

The results demonstrate that GLM 5.2 is viable for local execution on Apple Silicon hardware, offering a functional alternative to other models like Qwen3.6 27B for specific coding tasks.