🚀 LLM Performance on Apple Silicon: Test Results
Inference speed measurements for various LLMs were conducted on an M4 Mac mini (16 GB, 120 GB/s) via Ollama 0.31.2. The Llama 3.2 3B model showed 46.7 tok/s, Mistral 7B — 22.8 tok/s, and DeepSeek R1 8B — 20 tok/s.
🌍 The data confirms a direct correlation between LLM generation speed on the Apple Silicon architecture and memory bandwidth, which is critical when planning local AI infrastructures.
👤 You can evaluate the real-world performance of specific models (Llama, Mistral, Qwen) on the M4 Mac mini to understand whether such devices are suitable for running neural networks locally.
Source 1: https://macyou.co/benchmarks