Benchmarks

Apple Silicon LLM inference benchmarks

Measured on our own production fleet — not estimated from spec sheets. Each figure is the median of 3 runs via the Ollama API with a fixed prompt and 512-token generation budget. Last updated 2026-07-13 · Ollama 0.31.2 · raw data (JSON, CC BY 4.0).

M4 (Mac mini)16 GB, 120 GB/s

ModelParamsQuantGeneration (tok/s)Prompt processing (tok/s)Guide
Llama 3.2 3B llama3.2:3b3BQ4_K_M46.7 ±0.11720
Mistral 7B mistral:7b7BQ4_K_M22.8 ±0.1645
Qwen 2.5 7B qwen2.5:7b7BQ4_K_M22.3 ±0.11130
Llama 3.1 8B llama3.1:8b8BQ4_K_M21.2 ±0.1587hardware guide →
DeepSeek R1 8B deepseek-r1:8b8BQ4_K_M20 ±0.1531
Qwen 2.5 14B qwen2.5:14b14BQ4_K_M11.7 ±0.1606

Coming next

The same suite will run on every chip in the fleet as capacity allows: M4 Pro (Mac mini) (273 GB/s), M4 Max (Mac Studio) (546 GB/s), M3 Ultra (Mac Studio) (819 GB/s). Generation speed on Apple Silicon is dominated by memory bandwidth, so expect roughly proportional scaling — but we publish only what we measure.

Methodology

Runtime: Ollama 0.31.2 (llama.cpp + Metal) on macOS 15.3.1 (Sequoia), default settings, launched fresh for the suite. Models pulled from the official Ollama library at their default quantization (Q4_K_M unless noted).

Each model gets one warm-up request (discarded), then 3recorded runs of the same prompt (“Write a 300-word explanation of how attention works in transformer models, aimed at a junior developer. Use one concrete analogy.”) with num_predict=512, temperature=0.7. Speeds come from Ollama's own counters: generation = eval_count / eval_duration, prompt processing = prompt_eval_count / prompt_eval_duration. We publish the median and the max−min spread.

Measured on production-identical fleet hardware. One warm-up request per model (discarded). Machine was otherwise idle apart from baseline system daemons.

The raw per-run records are in the JSON dataset (CC BY 4.0 — cite “Macyou Apple Silicon LLM Benchmarks” with a link). Spot an issue with the method? Tell us: [email protected].

Want these numbers as your endpoint?

Every model above deploys from the catalog in about 5 minutes with an OpenAI-compatible API, on hardware identical to the benchmark machines. Fixed monthly price, no per-token fees.