What is \"Quantization\" (like Q4_K_M or Q8)?
Quantization is essentially JPEG compression for AI models. Out of the box, models store their weights as high-precision 16-bit numbers (FP16). That makes them huge—an 8-billion-parameter model needs roughly 16GB of VRAM at full precision. Quantization rounds those numbers down into smaller 8-bit or 4-bit integers, shrinking the file by 50% to 75% without wrecking the model's smarts. Because the file is smaller, your GPU reads it faster and spits out text quicker.
When you download models in apps like LM Studio or Ollama, the suffix at the end of the filename tells you how aggressively it's been compressed:
| Label | Precision | 8B Model Size (incl. context) | Best For |
|---|---|---|---|
| Q4_K_M (Recommended) | 4-bit (Mixed) | ~5.5 – 6 GB | The sweet spot. Cuts size by ~70% with less than 1–3% quality loss, fitting comfortably on 8GB GPUs. |
| Q5_K_M / Q6_K | 5-bit / 6-bit | ~6.5 – 7.5 GB | A solid middle ground if you're running a 12GB–16GB GPU with VRAM to spare. |
| Q8_0 | 8-bit | ~9 – 10 GB | Virtually indistinguishable from the original unquantized model; worth grabbing if you have plenty of headroom. |
What does _K_M actually mean? In modern GGUF model files, K means the quantization method protects the model's most critical "thinking" layers at higher precision while squeezing less sensitive layers harder. The M stands for Medium.
If you have to choose, always run a larger model at Q4_K_M over a smaller model at Q8. A 14B model at Q4_K_M will noticeably outperform an 8B model at Q8 while eating up roughly the same amount of VRAM.
Looking for verified hardware rigs?
Check our real-time £/GB VRAM index to find systems optimized for local model inference.