Inference · compare

Gemma 3 270M vs Gemma 4 E2B.

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.

4machine kinds both ran on
MachineGemma 3 270M tok/sGemma 4 E2B tok/sContext A / B
Linux, 4 CPUs · 31 GB37Q4_K_M (4-bit) by unsloth84Q4_0 (4-bit) by unsloth— / 128K
L4 · 23 GB292Q4_K_M (4-bit) by unsloth150Q4_K_M (4-bit) by google— / 128K
GeForce RTX 3080 Ti · 12 GB286Q4_K_M (4-bit) by unsloth183IQ4_XS (4-bit) by unsloth— / 4K
Windows, 8 CPUs · 16 GB52Q4_K_M (4-bit) by unsloth13Q3_K_S (3-bit) by unsloth— / 4K