Ollama Just Got 93% Faster on Mac. Here's How to Enable It.

ollamaapple-siliconmlxlocal-llminference-speed

Abstraction: Ollama MLX backend delivers 93% decode speedup on Apple Silicon Macs

Key points:

Ollama 0.19 (March 31, 2026) ships MLX backend preview; MLX uses Apple Silicon's unified memory natively, eliminating the CPU-to-GPU data copies that Metal's discrete-GPU model incurs
Measured gains on M4 Max 64GB: decode 57.8 → 111.4 tokens/s (+93%), prefill 1,147 → 1,803 tokens/s (1.6×); total response time nearly halved
Requires 32GB+ unified memory; base M1/M2/M3/M4 (8/16GB) are excluded due to MLX memory management overhead
Enable via environment variable: OLLAMA_MLX=1 before starting the server
As of 0.19, only Qwen3.5 is supported by the MLX backend; unsupported models silently fall back to Metal with no speedup
Ollama 0.20 (in development) adds Gemma 4; team is implementing per-architecture MLX support incrementally

Connections: Ollama · Apple · Local LLM Inference · Apple Silicon · Mlx Framework