gpu4.pics
data via LocalScore
Apple M1
llamafile 0.9.2
8 GB unified Apple
running Llama 3.2 1B Instruct· Q4_K_M· 1.5B
80
tokens / sec
456 · Passable
Generation
80tok/s
higher is better
Prompt / prefill
1,192tok/s
higher is better
Time to first token
1.01 s
lower is better
Efficiency
3.54 tok/s/W
Electricity cost
$0.013/Mtok
Power draw
23 W
Apple M1 Pro 8P+2E · 32 GB RAM · Darwin · llamafile
gpu4.pics/r/ls-result-16

All benchmarks on this rig

ModelQuantGenPromptTTFTPowerPerf/W$/MtokScore
Llama 3.2 1B Instruct1.5BQ4_K_M801,1921.01 s23 W3.54 tok/s/W$0.013/Mtok456

How it stacks up — Llama 3.2 1B Instruct

Apple M1
80
NVIDIA L40S
355
NVIDIA RTX 3090 Ti
354
NVIDIA H100
335
NVIDIA RTX 3090
330
NVIDIA RTX A6000
315
NVIDIA A100 80GB
308