M4 Air beginner seeks local LLM help

Local inference on Apple Silicon is finally beginner-friendly, but 16GB still rewards restraint over brute force.

• –LM Studio looks like the smoothest on-ramp: official docs say it supports Apple Silicon, recommends 16GB+ RAM, and runs both `llama.cpp` and Apple MLX models.
• –Ollama is still a strong CLI/API option on Mac, but its docs also warn that model storage can quickly reach tens to hundreds of GB, so disk space and model choice matter early.
• –The practical starter zone is 4B-8B class models, with Qwen3 4B, Gemma 3 4B, and DeepSeek-R1 7B/8B distilled variants fitting this setup much better than 14B+.
• –The real beginner trap is optimizing for size first; local agents get useful when the workflow is simple enough to iterate on, not when the benchmark is impressive.