Apple M3 Ultra (80-core GPU) × Qwen2.5 14B Instruct
Qwen2.5 14B Instruct on Apple M3 Ultra (80-core GPU)
36.7best gen tok/s
590best prompt tok/s
4,352max context tested
6results
Measured · all numbers below come from the linked sources
llama.cpp
| Quant | Gen tok/s | Prompt tok/s | Source | Context tested | TTFT | Build |
|---|---|---|---|---|---|---|
| Q4_K - Medium | 36.7 | 568 | LocalScore ↗ | 512+128 | 2,077 ms | llamafile 0.9.2 |
| Q4_K - Medium | 35.9 | 590 | LocalScore ↗ | 4,352 | 2,013 ms | llamafile 0.9.2 |
| Q4_K - Medium | 35.9 | 579 | LocalScore ↗ | 512+128 | 2,033 ms | llamafile 0.9.2 |
| Q4_K - Medium | 35.9 | 568 | LocalScore ↗ | 512+128 | 2,088 ms | llamafile 0.9.2 |
| Q4_K - Medium | 35.8 | 583 | LocalScore ↗ | 512+128 | 2,040 ms | llamafile 0.9.2 |
| Q4_K - Medium | 35.7 | 580 | LocalScore ↗ | 512+128 | 2,043 ms | llamafile 0.9.2 |
Ollama
No measured Ollama results yet. Submit yours →
MLX
No measured MLX results yet. Submit yours →
"Context tested" is the largest prompt+generation length actually benchmarked (tokens), not a memory-fit estimate.
What else can Apple M3 Ultra (80-core GPU) run? →
Includes results from LocalScore, a Mozilla Builders project by cjpais, mirrored in the cn0303/local-llm-speed-benchmarks dataset. LocalScore runs use llamafile, which is built on llama.cpp.