M4 vs M3

Side-by-side LLM inference benchmarks: M4 versus M3 across 3 models. Evidence-backed tok/s measurements with confidence metadata.

3Shared models

M4Wins 3 of 3

30%Avg speed advantage

6Measurements used

M4 is faster in 3 of 3 models tested. Average advantage: 30%.

Model-by-model comparison

Each row shows the fastest published generation speed for that model on each chip family. Higher tok/s is better. Evidence badges show data provenance.

Model	M4	M3	Difference	Evidence
llama-3-1-8b-instruct	16.8 tok/s Q4_K_M	13.5 tok/s Q4_K_M	25% M4	CommunityCommunity
llama-3-2-1b-instruct	76.2 tok/s Q4_K_M	67.2 tok/s Q4_K_M	13% M4	CommunityCommunity
qwen-2-5-14b-instruct	9.2 tok/s Q4_K_M	6.1 tok/s Q4_K_M	51% M4	CommunityCommunity

Data confidence

This comparison uses 6 measurements. 6 are community-reported.

All numbers reflect generation speed (tok/s) at the best available quantization for each side. Quantization levels may differ between families. Where quant levels differ, the comparison shows each chip at its measured best — not a controlled variable.

Chip variants in this comparison

M4 10 core gpu M4 8 core gpu M4 gpu Count Not Published

M3 10 core gpu M3 gpu Count Not Published

Data

benchmarks.json — full dataset · benchmarks.csv — CSV export

See all chips →