How much memory and disk space do offline AI models need?
A 4 billion parameter model needs about 2.5 GB of disk and roughly about 4 GB of memory while it answers, so it suits 8 GB laptops. An 8 billion parameter model needs about 5.0 GB of disk and roughly about 8 GB of memory, right for 16 GB. A 14 billion parameter model needs about 9.0 GB of disk and roughly about 15 GB of memory, so plan on 32 GB. Keep a few gigabytes of disk free on top.
Two kinds of space matter. Disk space (storage) holds the model file once it is downloaded. Memory (RAM) is where the model sits while it is answering. Disk is easy to free up. Memory is fixed when you buy the laptop, so it is the number that decides which model you can run.
The simple table
| Model size | Example | Download (disk) | Memory while answering | Laptop memory to aim for |
|---|---|---|---|---|
| About 4 billion parameters | Qwen3 4B Instruct 2507 | 2.5 GB | about 4 GB | 8 GB |
| About 8 billion parameters | Qwen3 8B | 5.0 GB | about 8 GB | 16 GB |
| About 14 billion parameters | Qwen3 14B | 9.0 GB | about 15 GB | 32 GB, or a graphics card with 12 GB+ |
| Search helper | nomic-embed-text v1.5 | 146 MB | Small | Any |
Downloads are the 4-bit (Q4_K_M) GGUF files published on Hugging Face (see Sources), rounded. Memory while answering is our estimate and grows with how much text the model is working on at once. The search helper is a tiny model that finds the right passages in your files.
Why memory matters more than disk
While the model writes, the whole file needs to be in memory, plus some working memory for the conversation and any document it is reading. Your operating system, browser and other apps need memory at the same time. That is why the table suggests a laptop with roughly twice the memory the model itself uses. If memory runs short, the computer starts swapping to disk and answers slow right down.
On Apple Silicon Macs, the graphics chip shares the main memory, which is a big advantage for AI. On Windows laptops, a graphics card with its own memory can hold the model instead. More in the Mac and Windows guides.
Why the numbers look different on your computer
File sizes on Hugging Face count a gigabyte as 1,000,000,000 bytes. Windows counts in binary units of 1,073,741,824 bytes but still calls them GB, so a 2.5 GB download shows as about 2.3 GB in File Explorer and a 9.0 GB file as about 8.4 GB. macOS uses the same decimal units as Hugging Face. Nothing is missing; it is just a different ruler.
Check your laptop
- Mac memory: Apple menu, About This Mac, Memory.
- Mac disk: Apple menu, System Settings, General, Storage.
- Windows memory: Settings, System, About, Installed RAM.
- Windows disk: File Explorer, This PC, free space on the C: drive.
Short on space?
- Empty the Bin (Mac) or Recycle Bin (Windows).
- Clear your Downloads folder of old installers and duplicate files.
- Move old photos and videos to an external drive.
- If disk is still tight, choose the smaller model. It is quicker to download too.
How WorkOffline handles this
WorkOffline checks memory and free disk on first run and recommends a size that fits. If there is not enough free disk for that size, it steps down to a smaller one and tells you why, and if there is not enough even for Lite, it tells you how much to free up. The model is downloaded once, checked against its published fingerprint (a SHA-256 hash) before use, and then works with no connection.
Every ticket includes chat with your files, writing help and compress for email, from one flat white. See tickets and prices.
Frequently asked questions
How much RAM do I need to run AI locally?
8 GB is enough for a small model of about 4 billion parameters. 16 GB runs an 8 billion parameter model comfortably. For a 14 billion parameter model, plan on 32 GB, or a graphics card with 12 GB or more of its own memory.
How much disk space does an offline AI model take?
About 2.5 GB for a 4 billion parameter model, 5.0 GB for 8 billion and 9.0 GB for 14 billion, in the common 4-bit format. Keep a few gigabytes free on top for your system.
Can I store the AI model on an external drive?
Many tools allow it, but the model loads more slowly from a USB drive and stops working when the drive is unplugged. The laptop's own drive is the better place.
Does the model use memory when I am not using it?
It depends on the app. A model uses memory while it is loaded. Some apps keep it loaded for quick replies and release it later.
Why does my 16 GB laptop say 15.6 GB?
Some memory is reserved for the system or the built-in graphics, and some tools count in binary units. Both are normal.
Sources
- Qwen3-4B-Instruct-2507 model card (Hugging Face): 4.0B parameters, Apache 2.0
- Qwen3-4B-Instruct-2507 GGUF files (Unsloth, Hugging Face): Q4_K_M file 2,497,281,120 bytes
- Qwen3-8B model card (Hugging Face): 8.2B parameters, Apache 2.0
- Qwen3-8B-GGUF (Qwen, Hugging Face): Q4_K_M file 5,027,783,488 bytes
- Qwen3-14B model card (Hugging Face): 14.8B parameters, Apache 2.0
- Qwen3-14B-GGUF (Qwen, Hugging Face): Q4_K_M file 9,001,752,960 bytes
- nomic-embed-text-v1.5 GGUF (Nomic, Hugging Face): Q8_0 file 146,146,432 bytes, Apache 2.0
- [Apple Support: find out how much memory your Mac has (About This Mac)]
- [Microsoft Support: find your PC specs in Settings, System, About]
An evergreen guide from WorkOffline. No tracking scripts on this page.