Skip to content
Local AI Wizard
Your machine
Theme

How much memory do I need to run an LLM?

Model weights, working room, the conversation and your system, added up.

Memory · updated

Short answer

~29GB

About 29 GB for a 27B model in the usual Q4 file. Rule of thumb: 0.625 GB per billion parameters, plus room to work.

  • Weights: 16.9 GB for 27B in Q4.
  • Working room: +20% while it runs.
  • System reserve: 8 GB on a 32 GB+ Mac or PC, less on smaller machines.
  • Graphics card: About 23 GB of VRAM for the same model.

Planning estimate with the site's formula, not a measurement. Methodology

The formula

A model runs from memory: every parameter has to be loaded. llama.cpp states it plainly: memory and disk requirements are the same (llama.cpp · quantize, read Oct 17, 2026). On top of the file, the site adds three parts:

memory ≈ weights + 20% working room + long-conversation budget + system reserve

Weights depend on the file version: Q4 stores about 0.625 byte per parameter, Q8 about 1.125 and FP16 2 (what Q4 means). The reserve is memory the computer keeps for itself; on a graphics card the system uses its own RAM, so the reserve is smaller (VRAM and unified memory).

By model size

Q4 file, short conversations, on a Mac or PC with 32 GB or more (8 GB reserve); machines under 32 GB reserve 25% of their memory (2 to 8 GB), so they need less, as the machines below show. Estimates rounded up to whole GB.

4B parameters
~11 GB on a 32 GB+ Mac or PC~5 GB on a graphics card · Q4 file ~2.5 GB
9B parameters
~15 GB on a 32 GB+ Mac or PC~9 GB on a graphics card · Q4 file ~5.6 GB
27B parameters
~29 GB on a 32 GB+ Mac or PC~23 GB on a graphics card · Q4 file ~16.9 GB
70B parameters
~61 GB on a 32 GB+ Mac or PC~55 GB on a graphics card · Q4 file ~43.8 GB

Check with a real file

Qwen3.6 27B: published file sizes next to the site's weight estimate.

Qwen3.6 27B files on Hugging Face, read Oct 16, 2026. GB = 10⁹ bytes.
FilePublished size (GB)Site weight estimate (GB)
Q4_K_M16.816.9
Q8_028.630.4
BF1653.854

The estimate stays within 7% of the real files, so the formula is a fair way to plan before you download. Real files differ a little because each format stores its own scale factors and keeps some parts of the model at other precisions. (unsloth/Qwen3.6-27B-GGUF, read Oct 16, 2026)

The same holds at larger sizes: llama.cpp lists Llama 3.1 70B at 280.9 GB in its original file and 43.1 GB in Q4_K_M; the site's Q4 weight estimate for 70B is 43.8 GB. (llama.cpp · quantize, read Oct 17, 2026)

On real machines

The largest of the sizes above whose estimate, with this machine's own reserve, stays within the memory it offers to a model.

Mac mini M4 · 16 GB unified
Up to 9B in Q4 (~11 GB)Offers 16 GB to a model
Mac mini M4 Pro · 48 GB unified
Up to 27B in Q4 (~29 GB)Offers 44 GB to a model
RTX 4060 · 8 GB VRAM
Up to 4B in Q4 (~5 GB)Offers 8 GB to a model
RTX 4090 · 24 GB VRAM
Up to 27B in Q4 (~23 GB)Offers 24 GB to a model

Calculator

Change the size, the file version, the conversation and the machine. Planning estimate, not a test; it never estimates speed.

File version (quantization)

About 0.6 byte per parameter: the usual local download.

Long-conversation budget

Short chats fit in the working room. Long documents and long chats need extra memory for the conversation.

Machine

~20.3 GBneeded, without the system reserve

Choose a machine to compare. Without one, the system reserve is left out: it depends on the machine.

  • Model weights16.9 GB
  • Working room (20%)3.4 GB
  • Long-conversation budget0 GB
  • System reserveDepends on the machine: choose one