A performance report details the Gemma-4-31B model's capabilities on Cloud TPU v6e-4 hardware, achieving a peak prefill throughput of 463,345 tokens/sec. The benchmarks indicate that the dense 31B model offers comparable throughput to a 26B MoE model on the same hardware, with better latency for interactive tasks. However, the MoE model demonstrates superior compute efficiency and can handle much larger context windows. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT Demonstrates hardware-software co-optimization for dense models, offering insights into performance trade-offs against MoE architectures.
RANK_REASON This is a performance report and benchmark analysis of a specific model on particular hardware, not a new model release or significant industry event. [lever_c_demoted from research: ic=1 ai=1.0]