Builds / Apple M3 Pro (unified)
Apple M3 Pro (18 GB): Llama 3.1 8B at 27.2 tok/s, Qwen3 14B at 14.5 (Ollama, median of 6 runs)
Last updated:
Source: localaimaster.com ↗📏 How measured: reported by the original source; RigBase re-checked the numbers at that source
Legacy models only. Every model measured on this rig now has a newer successor: Llama 3.2 3B → Muse Glimmer 30B, Llama 3.1 8B → Muse Glimmer 30B, DeepSeek R1 Distill 14B → DeepSeek V4.1 Flash. The hardware numbers are still a useful reference.
- GPU(s)
- Apple M3 Pro (unified)
- VRAM / unified memory (GB)
- 18
- CPU
- Apple M3 Pro
- Motherboard
- not stated
- RAM
- 18 GB unified memory (150 GB/s)
- PSU
- not stated
- Other
- macOS; Ollama 0.40.13; the publisher's own machine
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Llama 3.2 3B legacy | Q4_K_M (2.02 GB, Ollama tag llama3.2:3b) | Ollama 0.40.13 (Metal) | 55.6 | not stated | context: short prompt, up to 512 output tokens📏 Local AI Master harness (benchmark-ollama.py): same ~200-word technical prompt, output capped at 512 tokens, fixed seed; 7 runs, first discarded as warm-up, median of the 6 valid runs; rate from Ollama's eval_count/eval_duration; Ollama 0.40.13, measured 2 Nov 2026; slowest-fastest valid run 53.4-55.8; 55.3 on a second sweep the same day |
| Llama 3.1 8B legacy | Q4_K_M (4.92 GB, Ollama tag llama3.1:8b) | Ollama 0.40.13 (Metal) | 27.2 | not stated | context: short prompt, up to 512 output tokens📏 Local AI Master harness (benchmark-ollama.py): same ~200-word technical prompt, output capped at 512 tokens, fixed seed; 7 runs, first discarded as warm-up, median of the 6 valid runs; rate from Ollama's eval_count/eval_duration; Ollama 0.40.13, measured 2 Nov 2026; slowest-fastest valid run 26.9-27.4; 26.8 on a second sweep the same day; 89% of the bandwidth ceiling |
| Qwen3 8B legacy | Q4_K_M (5.23 GB, Ollama tag qwen3:8b) | Ollama 0.40.13 (Metal) | 25.3 | not stated | context: short prompt, up to 512 output tokens📏 Local AI Master harness (benchmark-ollama.py): same ~200-word technical prompt, output capped at 512 tokens, fixed seed; 7 runs, first discarded as warm-up, median of the 6 valid runs; rate from Ollama's eval_count/eval_duration; Ollama 0.40.13, measured 2 Nov 2026; slowest-fastest valid run 25.2-25.3; thinking model, timed over every emitted token including reasoning |
| Gemma 3 12B legacy | Q4_K_M (7.31 GB, Ollama tag gemma3:12b) | Ollama 0.40.13 (Metal) | 16.7 | not stated | context: short prompt, up to 512 output tokens📏 Local AI Master harness (benchmark-ollama.py): same ~200-word technical prompt, output capped at 512 tokens, fixed seed; 7 runs, first discarded as warm-up, median of the 6 valid runs; rate from Ollama's eval_count/eval_duration; Ollama 0.40.13, measured 2 Nov 2026; slowest-fastest valid run 16.1-16.8 |
| DeepSeek R1 Distill 14B legacy | Q4_K_M (8.99 GB, Ollama tag deepseek-r1:14b) | Ollama 0.40.13 (Metal) | 14.7 | not stated | context: short prompt, up to 512 output tokens📏 Local AI Master harness (benchmark-ollama.py): same ~200-word technical prompt, output capped at 512 tokens, fixed seed; 7 runs, first discarded as warm-up, median of the 6 valid runs; rate from Ollama's eval_count/eval_duration; Ollama 0.40.13, measured 2 Nov 2026; slowest-fastest valid run 14.6-14.8; thinking model, timed over every emitted token including reasoning |
| Qwen3 14B legacy | Q4_K_M (9.28 GB, Ollama tag qwen3:14b) | Ollama 0.40.13 (Metal) | 14.5 | not stated | context: short prompt, up to 512 output tokens📏 Local AI Master harness (benchmark-ollama.py): same ~200-word technical prompt, output capped at 512 tokens, fixed seed; 7 runs, first discarded as warm-up, median of the 6 valid runs; rate from Ollama's eval_count/eval_duration; Ollama 0.40.13, measured 2 Nov 2026; slowest-fastest valid run 14.4-14.5; thinking model, timed over every emitted token including reasoning; 90% of the bandwidth ceiling |
Local AI Master's open benchmark table (CC BY 4.0), published 2 Nov 2026, measured on the publisher's own M3 Pro 18 GB. Every row passes their validity gate (median must stay under 1.05x memory bandwidth / file size). They held back Qwen3.5 on Apple Silicon (Ollama served an MLX build with multi-token prediction) and Qwen3.6 27B everywhere (MTP), so those are not listed. Their NVIDIA L4 rows are a rented cloud VM and are not listed. Generation speed only; no prompt-processing figures.
Source: https://localaimaster.com/benchmarks
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Embed this result
Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds — numbers stay on the build page.
Markdown
[](https://getrigbase.com/builds/macbook-pro-m3-pro-18gb-ollama-localaimaster/)
HTML
<a href="https://getrigbase.com/builds/macbook-pro-m3-pro-18gb-ollama-localaimaster/" rel="noopener"><img src="https://getrigbase.com/badge/macbook-pro-m3-pro-18gb-ollama-localaimaster.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)