GLM-5.3-MLX

Overview

GLM-5.3 quantized for Apple Silicon and the MLX CUDA backend with OrcaSAQ sensitivity-aware mixed precision - attention pinned at 8-bit and near-lossless output at 4-bit (cosine 0.9968 vs FP8). The repo root is the recommended 4-bit build; needs mlx-vlm 0.6.17+.

Good for

• agentic
• chat
• coding

Recommended hardware

≈ 512 GB VRAM to run this 753B model at the OrcaSAQ-4bit quantization.

Needs a multi-GPU server node (e.g. 8x 80 GB H100, or 4x 192 GB)

Links

• View on HuggingFace
• Run on Infersec