Google says Gemini 4 Argon speeds up Gemma 4 E2B inference 2.85x on A100

Google says its Gemini 4 Argon optimization raised inference speed for the Gemma 4 E2B model on an A100 GPU running vLLM, from about 37,900 to over 108,000 tokens per second, a 2.85x speedup.

Google reports 100% token parity and no quality drop, and says the gain came in a few days of optimization work. As a demonstration, it says the model generated 64 live Python simulations at once.

The figures come from a Google post on X. The post does not include a benchmark setup or independent verification.