M1 MacBook Air strains local LLMs

16GB gets you into local AI, but it’s the floor, not the sweet spot, once you stack a chat model, retrieval, and another app on top.

• –Apple’s 16GB M1 Air can run 7B/8B-class quantized models, and Ollama’s packaged Llama 3.1 8B and Qwen2.5 7B builds are both around 5GB, but that still leaves limited headroom for long contexts and background processes.
• –Best default picks here are Llama 3.1 8B Instruct, Qwen2.5 7B Instruct, and Mistral 7B; Qwen2.5 and Llama 3.1 both support 128K contexts, while Mistral stays especially light and stable.
• –If you want a NotebookLM-like local setup, think RAG stack first and model second: embeddings, indexing, and the UI all consume RAM too.
• –32GB is the practical minimum for a smoother experience, while 64GB is the comfort tier if you want bigger models, longer sessions, and fewer compromises.