Quantization: Fitting LLMs into VRAM

Quantization reduces the precision of model weights (from 16-bit to 8-bit or 4-bit), drastically reducing VRAM requirements with minimal accuracy loss.

Home / Guides / LLM Quantization Methods

GGUF (llama.cpp)

The standard for running models on CPU or Apple Silicon. It allows mixed-precision and offloading layers to the GPU, making it the most flexible format for local experimentation.

AWQ and GPTQ

Formats optimized for pure GPU inference. AWQ (Activation-aware Weight Quantization) generally offers better accuracy retention than GPTQ at 4-bit precision and is highly recommended for production GPU deployments (e.g., via vLLM).

Internal Resources