I Tried Local LLMs for Coding on My M3 MacBook Pro: What Worked and What Didn't

Описание и применение

Локальные модели LLM на MacBook Pro M3 могут эффективно выполнять небольшие задачи кода, но их задержка делает их непрактичными для средних проектов, где требуется агентный цикл. Лучший результат показал Gemma 4 31B‑MLX, однако общая скорость оставалась ниже, чем у облачных ассистентов.

Задача

Can a local LLM become a practical everyday coding assistant on a laptop?

Результат

Local LLMs on an M3 MacBook Pro can serve as a useful coding assistant for small, focused tasks, but their latency in agentic, repository‑wide workflows limits productivity for medium‑complexity projects. The best local experience was with Gemma 4 31B‑MLX, yet overall speed remained insufficient compared to cloud assistants.

Рабочий процесс

• Set up Ollama on M3 MacBook Pro and bind to 0.0.0.0:11434
• Configure VS Code on Lenovo to point to Mac’s IP
• Test network connectivity with curl to /api/tags
• Run and switch models with ollama run <model>
• Adjust context length via OLLAMA_CONTEXT_LENGTH
• Evaluate models (Gemma 4 26B, Gemma 4 26B‑MLX, Qwen3‑Coder 30B, Gemma 4 31B‑MLX) for speed and quality
• Use VS Code to ask for explanations, code snippets, refactorings, tests, design ideas
• Attempt agentic workflow: repository inspection, plan generation, edits, build/test feedback
• Measure latency per step and overall loop
• Compare local performance to cloud Copilot
• Decide on tool usage based on workflow needs

Доказательства

Gemma 4 26B worked well on my machine and was capable enough to make local coding feel practical.

Источник

For a medium‑complexity project with repeated planning, repository exploration, code changes, and validation, it was too slow for the way I wanted to work.

Источник