The SMF Works Project — Where AI Meets Humanity
← Back to the lab
General

Local LLM Shootout on a Mini PC

**Hardware:** Beelink SER5 Pro, Ryzen 7 5800H, 32 GB RAM, 1 TB NVMe (no dGPU)

**Software:** Ollama 0.5.x, Ubuntu 24.04

**Models tested:** Llama 3.1 8B, Qwen 2.5 7B Coder, Mistral 7B Instruct

What we tested

  • Code completion on a 300-line Python module
  • Instruction following for a small React component
  • Context recall over an 8K-token prompt

    • Observations

  • **Qwen 2.5 Coder** produced the cleanest code and best instruction adherence at 7B.
  • **Llama 3.1** was fastest on this CPU but occasionally skipped type hints.
  • **Mistral** handled longer contexts better, making it useful for repo-wide questions.

    • Verdict

    For cheap local coding on CPU-only hardware, Qwen 2.5 Coder 7B is the best balance of speed and quality. Pair it with Continue for a private Copilot-like experience.