Apple’s M5 Macs With macOS 26.2 Deliver Up to 4x Faster Local LLM Start Times in MLX

Overview

• Apple’s official benchmarks show up to roughly 4x improvements in time to first token versus M4 when running LLMs with MLX on M5 hardware.
• Dense 14B models drop below about 10 seconds to first token, while a 30B Mixture‑of‑Experts example falls under about 3 seconds.
• Subsequent‑token generation improves by 19–27% due to higher memory bandwidth on M5 Macs, measured at about 153GB/s versus 120GB/s on M4.
• Independent reporting cites a Qwen3‑14B 4‑bit run moving from roughly 36 seconds to around 8 seconds on M5, corroborating Apple’s results.
• MLX remains open source with Python, Swift, and C/C++ support, adds fast built‑in quantization for Hugging Face models, and requires macOS 26.2 to tap the M5 accelerators.