Quick verdict
Start with Qwen3 4B at 4-bit quantization and 4096-token context on a 16GB Windows laptop without a discrete GPU. For an 8GB machine, use a smaller 1B–2B model first. Expect CPU and memory bandwidth to limit speed.
What this profile is for
This profile assumes 16GB system RAM and no dedicated GPU memory. Check Settings → System → About for your installed RAM, and Task Manager → Performance for the GPU. “Shared GPU memory” borrows system RAM; it does not turn this laptop into an 8GB discrete-GPU machine.
Get the first model working
Follow install Ollama, then run ollama run qwen3:4b. Its Q4_K_M download is about 2.5GB. In the chat prompt enter /set parameter num_ctx 4096 and ask for a short rewrite. On a smaller machine, qwen3:1.7b is a lighter starting option.
If you prefer buttons, use install LM Studio, check its Windows CPU requirements, and select a small Qwen3 GGUF file. A supported integrated-GPU backend may help on some devices, but this starting plan does not depend on it. An NPU advertised on the laptop does not establish support for an arbitrary downloaded model.
Worked memory example
For a 16GB laptop, assume 2.5GB of model weights, a 1.5GB planning allowance for runtime and short context, and 6GB for Windows and ordinary applications. The total is 10GB, leaving approximately 6GB for growth. These are rounded budgeting assumptions; inspect actual memory use because the backend and context change the runtime cost.
Use the memory calculator with your real RAM and a small text-model configuration. Then watch Task Manager during a short chat. A model that fits but takes too long to answer has a speed problem; adding storage or stretching the context will not fix CPU throughput.
Decide whether to move up
Try a 7B–8B 4-bit model only after a small model answers at an acceptable pace and enough RAM remains. Close a browser-heavy workload before comparing. Test one representative question, then a normal conversation; do not start with an entire document library.
If a small model is still too slow, stay with short tasks or use another computer with a supported GPU. On most laptops a discrete GPU cannot simply be added internally.
Recommended next step
Follow the Windows setup comparison, or use slow-model troubleshooting when loading works but responses are impractical.