Open-Source AI Models Evaluated Against Google's Gemma on TPU Hardware

A benchmark tested small open-license AI models on a single Google Cloud TPU v5e chip, using Gemma 4 E2B as a baseline for performance. The evaluation measured how efficiently models like Qwen3-4B, SmolLM3-3B, and Granite 3.3-2B matched Gemma's accuracy and tool-calling abilities. Gemma 4 E2B maintained an advantage, particularly on tasks involving long prompts. The test aimed to identify capable, freely distributable models that could run on constrained, cost-effective hardware.