Skip to content
Bitpute
ToolsModelsGPUsCloudLearn
Learn

Learn

By Bitpute · Updated 24 July 2026 · How we estimate · Sources · Editorial policy · Version history · Report an error

The concepts behind every number on this site, in plain English — no math degree required.

FP16 vs Q8 vs Q6 vs Q4: which quantization should you use?

A practical comparison of FP16, Q8_0, Q6_K, Q5_K_M, Q4_K_M and Q4_0 — bits per weight, real VRAM for a 14B model, quality, and which to pick for your

How much system RAM do you need for local AI?

System RAM vs VRAM for running LLMs: 16GB minimum, 32GB comfortable, 64GB+ for offloading big models. How offload budget works, the Apple unified-memo

What is KV cache? (why long conversations eat your VRAM)

The KV cache stores attention keys and values for every token in context, and grows linearly with context length — at long context it can exceed the m

What is quantization? (and why it lets you run big models on small GPUs)

Quantization stores model weights in fewer bits to cut VRAM without retraining. What it does, how much you save, why Q4 isn't exactly 4 bits, and when

Why VRAM matters more than anything else for local AI

For local LLM inference, VRAM capacity decides what you can run and VRAM bandwidth decides how fast. Why memory, not compute, is the bottleneck — and

Want the formulas instead? See the methodology or the engineering notes.

Start here

GPU concepts for running LLMs maps how these pieces relate — which specification decides what, and in what order.

Core definitions

The two variables every calculation on this site depends on:

VRAMMemory bandwidthTraining memory