Inference · compare

Llama 3.2 1B vs Qwen 3 4B.

Side by side only on machines of the same kind, so the difference is the model, not the computer. Each figure is the fastest build of that model on that machine.

2machine kinds both ran on
MachineLlama 3.2 1B tok/sQwen 3 4B tok/sContext A / B
Windows, 8 CPUs · 16 GB18IQ3_XS (3-bit) by maziyarpanahi4.6IQ3_XXS (3-bit) by unsloth4K / 32K
Mac, 10 CPUs · 16 GB113Q4_K_M (4-bit) by maziyarpanahi23Q5_0 (5-bit) by Qwen8K / 4K