Quantization Explained: What You Actually Lose at Each Level

By , Ready Utilities

Last updated: September 2026

This page picks up where the hardware calculator's methodology leaves off: quantization explained in full, not just the formula. What does a model actually lose in quality as you drop from full precision down to the smallest usable size, and how do you pick the right point on that scale for what you're doing?

The full quantization ladder

Quantization stores a model's weights at lower numeric precision to shrink file size and memory use. "K-quants" (the _K_ formats below) are the current standard: they mix bit depths internally rather than using one uniform width, which is why a "4-bit" format averages closer to 4.5 bits per weight in practice.

FormatBits/weight7B sizeQuality
FP1616.0~13 GBFull precision, the inference baseline
Q8_0~8.0~6.7 GBNear-lossless
Q6_K~6.0~5.15 GBAlmost lossless
Q5_K_M~5.1~4.45 GBHigh quality, the safe balanced choice above Q4
Q4_K_M~4.5~3.80 GBMost popular default; good quality, meaningfully smaller
Q3_K_M~3.3~3.06 GBUsable but noticeably below FP16
Q2_K~2.6~2.9 GBSmallest, quality degrades significantly; last resort

Sizes scale roughly linearly with parameter count: a 32B model at a given format is about 4.6x the size shown above for 7B. Run the exact figure for your model through the hardware calculator rather than scaling by hand.

How much quality actually changes

Quantization loss is usually measured by perplexity, a score for how well a model predicts held-out text; a small increase in perplexity means the model got slightly worse at predicting text it wasn't trained on, which correlates with a general quality drop. The increase from FP16 stays close to zero through Q6_K and Q5_K_M, both considered near-lossless in practice. Q4_K_M is where the tradeoff becomes a real decision rather than a formality: still good quality, but a measurably bigger step down than anything above it. Below Q4, quality loss becomes progressively easier to notice in actual use, not just in a benchmark score.

Choosing the right level for your hardware

Work backward from your available memory rather than picking a format first: take your GPU's total VRAM, subtract roughly 500MB for system overhead, then subtract the memory your target context length needs (see the calculator's context setting for that figure). Whatever's left is your weight budget. Pick the highest quality format from the ladder above that fits inside it.

In practice, most people should treat Q5_K_M as the default when it fits, and drop to Q4_K_M when it doesn't; below Q4_K_M is worth it only when it's genuinely the difference between running a model and not running it at all, not as a routine choice. See every level's size side by side in the quantization calculator instead of doing this math by hand.

Legacy formats: Q4_0, Q5_0, and why to skip them

Older quantization formats without the _K_ naming (Q4_0, Q4_1, Q5_0, Q5_1) still show up in some model repositories. They use a simpler, uniform-precision approach that the K-quant formats have generally superseded on quality per gigabyte. Unless you're using a tool with a specific compatibility requirement for the legacy naming, pick the equivalent K-quant format instead.

The bits-per-weight, file-size, and perplexity-delta figures above are cross-checked against current GGUF quantization documentation and community benchmark write-ups (see the fact-check log). This page gets rechecked if a new quantization scheme becomes the community standard, since K-quants themselves superseded the legacy formats this way.

Sources

  1. Toni Sagrista (opens in new tab)
  2. Can It Run AI (opens in new tab)

Frequently asked questions

Is Q4_K_M good enough for everyday use?

Yes, for most people. It's the most widely used default for a reason: a meaningful size reduction from FP16 with quality that holds up for typical chat, writing, and coding use. Step up to Q5_K_M if you have the memory and want a safety margin.

When is it worth going below Q4?

Only when it's the difference between running a model at all and not running it. Quality loss below Q4_K_M becomes noticeable enough that a smaller model at Q4 often beats a larger model quantized down to Q2 or Q3.

Does quantization affect speed, not just quality?

Indirectly. A smaller quantized file usually loads faster and can run faster if it lets more of the model fit in fast GPU memory instead of spilling into slower system RAM. The quantization level itself doesn't change the underlying computation speed much.