Apple M1 Ultra (48-core GPU)

11measured results

139best gen tok/s*

128 GBreported memory

*Headline excludes tiny/test models (under 0.6B params or TinyStories-style). Open in rig checker →

llama.cpp

Model	Quant	Gen tok/s	Prompt tok/s	Source	Context tested	Build
Llama 3.2 1B Instruct	Q4_K - Medium	139	2,624	LocalScore ↗	512+128	llamafile 0.9.2	Details →
Llama 3.2 1B Instruct	Q4_K - Medium	138	2,537	LocalScore ↗	512+128	llamafile 0.9.2	Details →
Llama 3.2 1B Instruct	Q4_K - Medium	137	2,586	LocalScore ↗	512+128	llamafile 0.9.2	Details →
Llama 2 7B	Q4_0	74.9	772	llama.cpp (Metal) ↗	512+128	llama.cpp Metal	Details →
Llama 2 7B	Q8_0	55.7	783	llama.cpp (Metal) ↗	512+128	llama.cpp Metal	Details →
Meta Llama 3.1 8B Instruct	Q4_K - Medium	49.1	535	LocalScore ↗	512+128	llamafile 0.9.2	Details →
Meta Llama 3.1 8B Instruct	Q4_K - Medium	48.8	535	LocalScore ↗	512+128	llamafile 0.9.2	Details →
Meta Llama 3.1 8B Instruct	Q4_K - Medium	48.7	534	LocalScore ↗	512+128	llamafile 0.9.2	Details →
Llama 2 7B	F16	33.9	876	llama.cpp (Metal) ↗	512+128	llama.cpp Metal	Details →
Qwen2.5 14B Instruct	Q4_K - Medium	27.9	290	LocalScore ↗	512+128	llamafile 0.9.2	Details →
Qwen2.5 14B Instruct	Q4_K - Medium	27.7	289	LocalScore ↗	512+128	llamafile 0.9.2	Details →

Ollama

No measured Ollama results yet. Submit yours →

MLX

No measured MLX results yet. Submit yours →

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.