What happens if you go fully offline for a week with only a $400 laptop that has no discrete GPU? I installed Ollama, pulled a few 7B-14B models, and deleted all API keys. It worked far better than expected. This article documents the results and full setup steps.

💡 What You Will Learn

What happens if you go fully offline for a week with only a $400 laptop that has no discrete GPU? I installed Ollama, pulled a few 7B-14B models, and deleted all API keys. It worked far better than ex

📜 Table of Contents

• What worked
• What didn't
• Reproduce in 4 steps
• When local is the right call
• Verdict

I Ran AI Fully Offline on a $400 Laptop - It Was Surprisingly Usable

Setup: ThinkPad T14s (AMD Ryzen 7, 16GB RAM, no GPU), Ollama with Qwen2.5-7B and Llama-3.1-8B, continue.dev for code completion. Test: one week with zero cloud API calls.

What worked

• Code completion: 1-2s latency, close to Copilot, fully usable
• Summarization and brainstorming: solid
• Short-text rewriting: good enough

What didn't

• Reasoning beyond ~5K tokens: forgets context
• Creative writing: flat
• Mixed Chinese-English text: poor

Reproduce in 4 steps

1. Install Ollama: curl -fsSL https://ollama.com/install.sh | sh
2. Pull models: ollama pull qwen2.5:7b and ollama pull llama3.1:8b
3. Configure continue.dev in VS Code with an Ollama provider pointing to localhost:11434
4. Tune context length for 16GB RAM; use 7B for completion, 14B for harder chat.

When local is the right call

Good: code completion, summarization, private data, offline environments. Bad: complex reasoning, long-form creative writing, multilingual text - send those to cloud flagship models.

Verdict

70-80% of what the average developer uses AI for can be handled by a local 7-8B model on consumer hardware. The gap is closing fast.