By chip

Local LLMs on every Apple Silicon chip

Two numbers separate these chips for local inference: how much unified memory they can hold, which decides what runs at all, and how fast that memory is, which decides how quickly it runs. Everything else is secondary. Pick a chip for the full model by model breakdown, or size a specific case in the calculator.

M6

153 GB/s

Up to 32 GB · 26.2 tok/s on an 8B

Tops out at Qwen 2.5 32B in Q4. In our lineup, on the waitlist.

M5 Ultra

1200 GB/s

Up to 512 GB · 130 tok/s on an 8B

Tops out at DeepSeek V3 in Q4. In our lineup, on the waitlist.

M5 Max (614 GB/s)

614 GB/s

Up to 128 GB · 84 tok/s on an 8B

Tops out at Mistral Medium 3.5 in Q4. In our lineup, on the waitlist.

M5 Pro

307 GB/s

Up to 64 GB · 48.4 tok/s on an 8B

Tops out at Qwen 2.5 72B in Q4. In our lineup, on the waitlist.

M4 Max (40-core GPU)

546 GB/s

Up to 128 GB · 77 tok/s on an 8B

Tops out at Mistral Medium 3.5 in Q4. We rent this chip.

M4 Pro

273 GB/s

Up to 64 GB · 43.8 tok/s on an 8B

Tops out at Qwen 2.5 72B in Q4. We rent this chip.

M4

120 GB/s

Up to 32 GB · 20.9 tok/s on an 8B

Tops out at Qwen 2.5 32B in Q4. We rent this chip.

M3 Ultra

819 GB/s

Up to 512 GB · 102 tok/s on an 8B

Tops out at DeepSeek V3 in Q4. We rent this chip.

M3 Max (40-core GPU)

400 GB/s

Up to 128 GB · 60 tok/s on an 8B

Tops out at Mistral Medium 3.5 in Q4.

M2 Ultra

800 GB/s

Up to 192 GB · 101 tok/s on an 8B

Tops out at Mistral Medium 3.5 in Q4.

M1 Max

400 GB/s

Up to 64 GB · 60 tok/s on an 8B

Tops out at Qwen 2.5 72B in Q4.

The app you run them in matters less than the memory: Ollama for an API, LM Studio for a desktop chat window, MLX for the most Apple native path. For the largest open models, see Llama 3.1 405B on Apple Silicon.

Bandwidth and memory figures are Apple’s own published specifications, verified 2026-09-11. Speeds are estimates from a model fitted to benchmarks we measured ourselves; the method is on the calculator page and the raw data on benchmarks.