Start small
The fastest path is to choose a model that comfortably fits your machine. You can replace it later without rebuilding the controller.
| Machine | Good starting point | What to expect |
|---|---|---|
| 8 GB RAM, integrated graphics | qwen3.5:0.8b or another ~1B model | Basic chat, slower reasoning, very low memory pressure |
| 16 GB RAM | qwen3.5:2b or qwen3.5:4b | Good tutorial experience on modern CPUs |
| 32 GB RAM / capable GPU | qwen3.5:9b | Stronger responses with more memory use |
| 64+ GB RAM / high-VRAM GPU | Larger 20B+ quantized models | Better capability, substantially larger downloads and memory needs |
RAM is not the whole story
Model size, quantization, context length, GPU/VRAM, memory bandwidth, and other running programs all affect speed. Treat these tiers as starting points, not guarantees.
Find your system information
Windows
Get-ComputerInfo | Select-Object CsTotalPhysicalMemory, CsProcessors, WindowsProductNameFor a quick GPU check, open Task Manager → Performance → GPU.
macOS
system_profiler SPHardwareDataType
system_profiler SPDisplaysDataTypeLinux
free -h
lscpu
lspci | grep -Ei "vga|3d|display"Choose a deployment style
Use native installation for the walkthrough. It gives Ollama and the speech components the most direct access to your hardware. A VM becomes attractive when isolation matters more than maximum performance.
Checkpoint
Write down your RAM amount and whether you have a dedicated GPU. Pick a small first model. You can always move up one size after the full assistant works.
Finished this chapter?
Your progress is stored only in this browser.