Which AI model should I use on a Windows laptop?
Pick by memory first. With 8 GB, use a model of about 4 billion parameters (about a 2.5 GB download). With 16 GB or more, an 8 billion parameter model (about 5.0 GB) is the best all-rounder. Only go up to a 14 billion parameter model (about 9.0 GB) if your laptop has a graphics card with 12 GB or more of its own memory. Without one, the bigger model fits but answers slowly.
Windows laptops vary much more than Macs. The same price can buy a thin laptop with 8 GB of memory and no graphics card, or a gaming laptop with 32 GB and a powerful one. So the right AI model depends on two checks: how much memory (RAM) you have, and whether you have a graphics card with its own memory.
Check your laptop
- Open Settings, then System, then About. Read Installed RAM.
- Press Ctrl, Shift and Esc together to open Task Manager. Choose Performance, then GPU.
- If you see Dedicated GPU memory with a figure such as 8 GB or 12 GB, that is your graphics card's own memory. If you only see shared memory, your graphics are built in.
The answer by memory and graphics card
| Your laptop | Model to use | Download | What to expect |
|---|---|---|---|
| 8 GB memory | About 4 billion parameters (Qwen3 4B Instruct 2507) | About 2.5 GB | Works for drafting, rewriting and questions about your files. Close heavy apps. |
| 16 GB memory | About 8 billion parameters (Qwen3 8B) | About 5.0 GB | The best all-rounder for most laptops. |
| 32 GB memory, built-in graphics | About 8 billion parameters | About 5.0 GB | Same model, more breathing room. The 14 billion model fits but replies slowly. |
| 16 GB or more, plus a graphics card with 12 GB or more of its own memory | About 14 billion parameters (Qwen3 14B) | About 9.0 GB | Better on harder questions, at a good pace. |
Download sizes are the 4-bit (Q4_K_M) files published on Hugging Face, see Sources. More detail in how much memory and disk space offline AI needs.
Why the graphics card matters
An AI model writes its answer a small piece at a time, and each piece means reading through the whole model in memory. Graphics cards are built to move lots of data quickly, so when the model fits in the card's own memory, answers flow much faster. That is why a gaming laptop with a 12 GB graphics card can run the 14 billion parameter model comfortably, while a thin laptop with 32 GB of ordinary memory can hold it but crawls through each answer.
If your graphics card has 6 or 8 GB of its own memory, it can still help with the smaller models. The rule stays simple: choose by main memory, and only step up to the large model when the graphics card can hold it.
Tips for Windows laptops
- Plug in for long jobs. Many Windows laptops slow the processor on battery. Plugged in, answers come faster.
- Check your power mode. In Settings, System, Power, "Best performance" helps when plugged in.
- Watch disk space. Windows updates need room too. Keep several gigabytes free beyond the model.
- Set expectations. A small model is a capable writing and document assistant, not an encyclopaedia. See what a small AI model is.
How WorkOffline handles this
On first run, WorkOffline checks your laptop's memory and free disk space, recommends Lite, Standard or Large, and explains why. With 32 GB but no large graphics card it suggests Standard to keep answers quick. The model downloads once on Wi-Fi, then everything runs on the laptop with no connection. Choosing and swapping models from the full model library is a Pro feature. See how WorkOffline picks a model.
Chat with your files, writing help and compress for email come with every ticket, from one flat white. See tickets and prices.
Frequently asked questions
Can I run AI offline on a Windows laptop with 8 GB of RAM?
Yes, with a small model of about 4 billion parameters. It uses roughly about 4 GB of memory while answering, so close big apps first. Answers are slower than on a newer laptop with more memory, but it works with no internet.
Do I need a graphics card (GPU) to run AI offline?
No. Small models run on the main processor. A graphics card with plenty of its own memory makes answers faster and makes the larger models practical, but it is not required.
What is "Dedicated GPU memory" in Task Manager?
It is the memory on the graphics card itself, separate from the laptop's main memory. Most thin laptops show only shared GPU memory, which means the graphics are built into the processor.
Does a Copilot+ PC or NPU help?
Some apps use the NPU (neural processing unit) on newer laptops, but many offline AI tools run on the processor and graphics instead. Memory is still the number to check first.
Is 32 GB of RAM worth it for offline AI?
It gives room for longer documents and other apps alongside the model. On its own, without a large graphics card, it does not make the bigger model fast, so the 8 billion parameter model is usually the better daily choice.
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
- [Microsoft Support: find your PC specs in Settings, System, About]
- [Microsoft Support: check GPU memory in Task Manager]
An evergreen guide from WorkOffline. No tracking scripts on this page.