Worker

Choosing a quantization

Quantization is how much each model weight is compressed when it's packed into a GGUF file. Lower precision means a smaller download and less memory to run, at a small cost to answer quality. It's the same model either way — only the numerical precision of its weights changes. When you install a catalogue model, most sizes let you choose a quantization; this page helps you pick.

The choices

Catalogue models offer up to three quantizations per size. They're listed smallest-to-largest in the console, and the console shows each option's exact download size before you install:

QuantizationQualitySize & memoryUse it when…
Q4_K_M (balanced default) Excellent — the standard choice for quantised open models Smallest You're not sure — start here. Best size-to-quality ratio.
Q6_K Near-lossless ~40–50% larger than Q4_K_M You have memory to spare and want a little more fidelity.
Q8_0 Effectively lossless vs. the full-precision model Roughly double Q4_K_M Quality matters most and the worker has plenty of VRAM/disk.

How to pick

  • Start with Q4_K_M. For almost every use case the quality difference from a higher quant is hard to notice, and it leaves the most memory free for a longer context window or more parallel requests.
  • Step up to Q6_K or Q8_0 if the worker has spare VRAM and you want the last few percent of fidelity — e.g. for code, math, or careful instruction-following where small errors compound.
  • Prefer a bigger model over a higher quant. If you have the memory, a larger model at Q4_K_M almost always beats a smaller model at Q8_0. Use the quant choice to fine-tune within a size you've already picked.

Whichever you choose, the worker downloads and checksum-verifies the GGUF into its models directory the same way — see Install a worker for the install flow. Not every size ships every quantization; the console only offers the ones that exist for that model.