Gemma 4 Ships with 1,200% Improvement in Agentic Tool Use: From 6.6% to 86.4% on τ2-bench Retail

Google Blog·medium signal

Google's Gemma 4 (released April 2, now trending via community GGUF benchmarks) shows the most dramatic agent capability jump in any open model: agentic tool use on τ2-bench Retail improved from 6.6% (Gemma 3 27B) to 86.4% (Gemma 4 31B) — a +1,200% change. The 26B MoE variant (3.8B active) hits Arena AI #6 while running on consumer GPUs. Math jumped +330% (AIME 2026: 20.8% → 89.2%) and coding +175% (LiveCodeBench: 29.1% → 80.0%). When in 'thinking' mode, the model produces 4,000+ tokens of reasoning before committing.