← Docs
Local AI updated July 3, 2026

What is quantization? Q4, Q8, and FP16 explained

Why the same model comes in five sizes, what the Q-numbers mean, and how much quality you actually lose.


The short version

A language model is billions of numbers (weights). Quantization stores those numbers with fewer bits each, making the model file smaller and faster to run — at a small cost in quality.

The original training format is usually FP16 (16 bits per weight). A 7-billion-parameter model at FP16 is about 14 GB. Quantized to 4 bits, the same model is roughly 4–5 GB and produces nearly the same output.

Reading the names

Formats you'll see on model downloads, from biggest to smallest:

Format Bits per weight (approx.) Quality
FP16 / BF16 16 The reference — what the lab trained
Q8_0 ~8.5 Effectively indistinguishable from FP16
Q6_K ~6.6 Excellent, hard to notice any loss
Q5_K_M ~5.7 Very good
Q4_K_M ~4.85 The community default — small, still strong
Q3_K_M ~3.9 Noticeable quality tradeoffs
Q2_K ~3.35 Desperate measures — often incoherent

The _K and _M suffixes refer to newer mixed-precision schemes that keep the most important weights at higher precision. In practice: Q4_K_M is the default for a reason — it roughly quarters the memory needed versus FP16 while keeping quality most people can't tell apart in normal use.

Why it matters for your hardware

Memory is the wall in local AI. Quantization is the lever that decides whether a model fits:

  • 7B model: 14 GB at FP16 → ~4.7 GB at Q4 — fits a 8–12 GB GPU
  • 32B model: 64 GB at FP16 → ~20 GB at Q4 — fits a 24 GB GPU
  • 70B model: 140 GB at FP16 → ~42 GB at Q4 — needs unified memory or multiple GPUs

Quantization also makes models faster: generation speed is limited by how quickly weights stream through memory, so a model half the size generates roughly twice as fast on the same hardware.

Rule of thumb

Start at Q4_K_M. If you have memory to spare, step up to Q5 or Q6 for a small quality bump. Only go below Q4 when it's the difference between running a model and not running it — a bigger model at Q4 usually beats a smaller model at Q8 anyway.

To see what fits your machine, try our Can I Run It? calculator — it does this math for any model, quant, and GPU combination.

Was this useful?

This doc is maintained by the humans who run BreadLab. Spotted an error? Tell us — we fix docs, not just typos.