Apple Silicon Benchmarks

Real speed measurements on Q4_K_M β€” tokens per second for popular models across all Apple Silicon chips.

πŸ‡·πŸ‡Ί Russian version: apple-silicon.ru.md

← Local models Β· Tools comparison

Sources: MLJourney, CraftRigs, MacYou, LLMCheck.

Chip 8B tok/s 14B tok/s 27–30B tok/s 70B tok/s
M1 8 GB 12–14
M1 16 GB 18–24 11–12
M2 16 GB 22–28 12–14
M2 Pro 32 GB 30–38 18–20 β€”
M2 Max 64 GB 50–65 30–35 β€” 10–12
M3 16 GB 22–33 12–15
M3 Pro 36 GB 35–39 20–22 12–15
M3 Max 128 GB 65–75 40–45 25–30 10–14
M4 16 GB 25–35 14–17
M4 Pro 48 GB 38–50 22–26 18–25 7–11
M4 Max 64 GB 65–85 40–50 28–35 12–18
M4 Max 128 GB 75–95 45–55 30–40 15–22
M4 Ultra 192 GB 100–140 60–80 40–60 25–35

= does not fit in RAM / constant swap (0.4–2.2 tok/s). All numbers are generation, not prefill. Prefill is typically 10–30Γ— faster. For NVIDIA/Windows/Linux performance will be significantly higher (RTX 4090 does 80–120 tok/s on 8B).

Speed by task (M1 16 GB, Q4_K_M)

Task Model tok/s Note
Autocomplete (FIM) Qwen 2.5 Coder 1.5B 30+
Main coder Qwen 2.5 Coder 7B 22–25
Fast chat Qwen 3.5 4B 28–35
Balanced chat Qwen 3.5 9B 10–13
Reasoning DeepSeek R1 Distill 14B 8–10

MLX vs llama.cpp on M4 Pro

M4 Pro 24 GB, Qwen3-Coder-30B MoE (source: asiai.dev):

Metric LM Studio (MLX) Ollama (llama.cpp) Difference
Throughput 102 tok/s 70 tok/s +46%
TTFT 291 ms 175 ms Ollama faster
GPU Power 12.4 W 15.4 W -20%
Process Memory 21.4 GB 41.6 GB -49%

MLX backend in Ollama

Since March 2026, Ollama can automatically switch to MLX backend on Mac with 32 GB+ RAM:

If you want Go to
Choose a model for your Mac models.md
Compare tools for max speed tools.md
How much memory do you need memory-and-context.md
Back to navigation README.md

In section: getting-started Β· running-models Β· models Β· catalog Β· quantization Β· memory-and-context Β· tools Β· advanced-setup Β· troubleshooting Β· apple-silicon
Related sections: Zero Level Β· AI Agents Β· Use Cases
Navigation: ← Local Models Β· ↑ Back to main Β· πŸ‡·πŸ‡Ί Русский