Inference · compare

Gemma 4 E2B vs Qwen2.5 0.5B Instruct.

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.

8machine kinds both ran on
MachineGemma 4 E2B tok/sQwen2.5 0.5B Instruct tok/sContext A / B
Linux, 4 CPUs · 31 GB84Q4_0 (4-bit) by unsloth239Q4_K_M (4-bit) by Qwen128K / 32K
GeForce RTX 3080 Ti · 12 GB183IQ4_XS (4-bit) by unsloth248Q4_K_M (4-bit) by Qwen4K / 8K
Linux PC with an NVIDIA GPU, 32 GB78Q4_K_M (4-bit) by unsloth239Q4_K_M (4-bit) by Qwen128K / 32K
Linux PC with no GPU, 32 GB25Q4_K_M (4-bit) by unsloth19Q4_K_M (4-bit) by Qwen4K / 4K
Windows PC with an NVIDIA GPU, 16 GB98Q4_K_M (4-bit) by unsloth248Q4_K_M (4-bit) by Qwen128K / 8K
Windows PC with an NVIDIA GPU, 8 GB15Q4_K_M (4-bit) by unsloth5.8Q4_K_M (4-bit) by Qwen— / —
Windows PC with Intel graphics, 16 GB6.8Q4_K_M (4-bit) by unsloth32Q4_K_M (4-bit) by Qwen— / 16K
Windows, 12 CPUs · 7 GB15Q4_K_M (4-bit) by unsloth5.8Q4_K_M (4-bit) by Qwen128K / 32K