Patrick Gawron
Date 2026-02-20 Read 1 min

Running local LLMs on a laptop

No data center needed. Modern laptop, 16-32 GB RAM, half-decent GPU → handles writing, code review, agent workflows.

Why local

→ Privacy - prompts never leave the machine. → Latency - no network round-trip, no rate limits, no cold starts. → Cost - zero per-token after setup. → Control - pick model, quantization, context window.

Hardware floor

ComponentMinimumComfortable
RAM16 GB32 GB+
GPU VRAM8 GB16 GB+
Disk20 GB100 GB+
OSAnyLinux/WSL

Runtimes (2026)

  1. llama.cpp / llama-server - fastest on CPU, GGUF quantizations, tiny footprint.
  2. Ollama - easiest onboarding, one-line model pulls, OpenAI-compatible API built in.
  3. LM Studio - GUI-first, great for non-CLI users.

→ All three speak OpenAI chat-completions API. Any GPT-4 client → works with local model.

Sweet-spot model

Early 2026: Qwen 3.5 14B at Q4_K_M.

  • ~9 GB on disk
  • ~10 GB VRAM
  • Capable for agent tasks
  • Fast enough you won’t watch it type
# Ollama example
ollama pull qwen3.5:14b
ollama run qwen3.5:14b "Write a haiku about SQLite."

Next steps

→ Point existing Claude Code / Cursor config at http://localhost:11434/v1. → Benchmark a few quants - Q4 usually best speed/quality trade-off. → See /experiments/local-agent-benchmarks/ for real numbers.

Lesson: the laptop already has enough silicon. The bottleneck is picking a small enough model and a runtime that gets out of the way.