Apple M4 (10-core GPU) × Meta Llama 3.1 8B Instruct
Meta Llama 3.1 8B Instruct on Apple M4 (10-core GPU)
18.1best gen tok/s
188best prompt tok/s
4,352max context tested
15results
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 | 18.1 | 185 | LocalScore ↗ | 512+128 | 6,903 ms | llamafile 0.9.2 |
| Q4_K - Medium | 17.8 | 182 | LocalScore ↗ | 4,352 | 6,996 ms | llamafile 0.9.2 |
| Q4_K - Medium | 17.7 | 186 | LocalScore ↗ | 512+128 | 6,851 ms | llamafile 0.9.2 |
| Q4_K - Medium | 17.7 | 185 | LocalScore ↗ | 512+128 | 6,890 ms | llamafile 0.9.2 |
| Q4_K - Medium | 17.6 | 188 | LocalScore ↗ | 512+128 | 6,851 ms | llamafile 0.9.2 |
| Q4_K - Medium | 17.4 | 180 | LocalScore ↗ | 4,352 | 6,987 ms | llamafile 0.9.2 |
| Q4_K - Medium | 17.1 | 178 | LocalScore ↗ | 4,352 | 7,146 ms | llamafile 0.9.2 |
| Q4_K - Medium | 16.9 | 173 | LocalScore ↗ | 512+128 | 7,441 ms | llamafile 0.9.2 |
| Q4_K - Medium | 16.6 | 152 | LocalScore ↗ | 512+128 | 8,114 ms | llamafile 0.9.2 |
| Q4_K - Medium | 16.6 | 159 | LocalScore ↗ | 512+128 | 7,930 ms | llamafile 0.9.2 |
| Q4_K - Medium | 16.4 | 172 | LocalScore ↗ | 4,352 | 7,421 ms | llamafile 0.9.2 |
| Q4_K - Medium | 15.9 | 146 | LocalScore ↗ | 512+128 | 8,259 ms | llamafile 0.9.2 |
| Q4_K - Medium | 14.7 | 136 | LocalScore ↗ | 512+128 | 8,667 ms | llamafile 0.9.2 |
| Q4_K - Medium | 13.0 | 120 | LocalScore ↗ | 4,352 | 9,783 ms | llamafile 0.9.2 |
| Q4_K - Medium | 11.7 | 121 | LocalScore ↗ | 512+128 | 10,739 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 M4 (10-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.