Best Ollama Models for Students on Apple Silicon in 2026

Ollama makes running open-weight models on Mac trivial. The question is which model to pull. The answer depends on your Mac's memory and your task. Here are the picks that actually work on Apple Silicon in 2026.

Ollama abstracts the model loading and quantization. `ollama pull mistral` gives you Mistral 7B in 4-bit quantization (about 4 GB on disk and in RAM). On an M1 with 8 GB unified memory, that's the comfortable ceiling. Move up to 16 GB and you can run Llama 3.1 8B alongside other apps. 32 GB unlocks Mixtral 8x7B (a mixture-of-experts model that punches above its weight) or a quantized Llama 70B. For coding, DeepSeek Coder 6.7B is purpose-trained on code and beats general 7B models on most LeetCode-style problems.

Download from ollama.com, run installer, then `ollama pull ` from Terminal. The installer also starts the local server.

Inference is GPU-heavy. A 7B model continuously generating drains roughly 10-15W on M-series. Plug in for long sessions.

Yes - configure the overlay URL to http://localhost:11434/ or use Ollama mode if your version supports it.