Google Gemma Inference Hits 108,000 Tokens per Second via Gemini 4 Argon

Google's optimization for the gemma-4-E2B-it model using the vLLM framework and Gemini 4 Argon increased throughput from approximately 37,900 to over 108,000 tokens per second on Nvidia A100 GPUs, a 2.85x improvement. The update was implemented across 8 pull requests and maintained 100% token parity with no drop in quality.