Local AI and Local Models107 · Module II · Lesson 05 of 11
Article · 12 min

GGUF, AWQ, GPTQ

The three formats, compared.

Summary

The three dominant quantization formats. GGUF (llama.cpp) is the ecosystem winner for CPU and Apple Silicon inference. AWQ (activation-aware) and GPTQ are optimized for GPU inference with vLLM, ExLlama, and similar. All three are real; the right choice depends on your runtime.

Objectives
  • 01Match each format to its runtime.
  • 02State the tradeoffs among the three.
  • 03Recognize the file extension of each.
The Lesson

GGUF

The llama.cpp format. Runs on CPU, GPU (CUDA/Metal/Vulkan), and Apple Silicon. Wide quantization variants (Q2_K through Q8_0, IQ variants). The de facto standard for local inference. Ollama, LM Studio, jan.ai all use it.

AWQ and GPTQ

Activation-aware Weight Quantization: preserves salient weights better; slight quality edge at 4-bit. GPTQ: older, still widely supported. Both are GPU-first. vLLM and TGI serve AWQ and GPTQ; llama.cpp does not natively. Choose based on your inference server, not the format's marketing.

Key Ideas
  • GGUF for local (llama.cpp ecosystem); AWQ/GPTQ for GPU serving.
  • The runtime determines the format, not the other way around.