MacBook Pro M5, 32GB wins for LLMs

The short answer is yes: 24GB can work, but 32GB gives materially better headroom once you add long context, the OS, VS Code, and background apps. For local coding assistants, the difference is less about raw model fit and more about avoiding swap and keeping the machine responsive.

• –Gemma 4 26B A4B is feasible on 24GB in lower-precision GGUF builds, but context length adds KV-cache overhead fast
• –Yellow memory pressure on macOS often means the system is already leaning on compression or swap, which hurts latency more than benchmark-style model fit
• –If the laptop is meant to be a daily coding machine, 32GB is the safer floor for sustained local inference
• –The model itself is only part of the load; Ollama plus VS Code, browser tabs, and extensions make 24GB feel tighter than the headline spec suggests
• –If the return window is open, this is one of the few cases where upgrading for memory, not CPU, is the pragmatic move