Local AI on cheap hardware: worth it?

Honest answer: a £600 machine with 16GB of RAM can run useful local models today — but only the small ones, and only if you’re patient. For privacy-sensitive notes, offline drafting, and tinkering, it’s genuinely worth it. For anything you need to be fast or frontier-quality, the cloud still wins on cost-per-result.

What actually runs on cheap hardware

With LM Studio or Ollama on a mid-range laptop:

• 7–8B models run fine for chat, summaries, and simple coding help.
• Quantized 13B models work but get sluggish.
• Anything bigger needs a dedicated GPU or a lot of patience.

The costs nobody mentions

The model is free. The rest isn’t:

• RAM is the real gate. 16GB is the floor; 32GB changes everything.
• Battery and heat. Local inference hammers both. This is a plugged-in activity.
• Your time. Setup, model selection, and prompt tuning add up.

When local genuinely beats the cloud

• You’re handling private data you can’t send anywhere.
• You’re offline or on a metered connection.
• You’re learning how these models actually behave.

When to just use the cloud

If you need GPT-class quality, fast responses, or you’re doing this occasionally, a few pounds of API credit will outperform a hardware upgrade every time.

Local AI on cheap hardware is real and improving fast. Just go in knowing it’s a hobby-grade setup, not a quiet free replacement for the frontier models.