Google Gemma 4: 31B Open Model Rivals Giants 20x Its Size on Math, Code, and Science

Gemma4All·medium signal

Google's Gemma 4 31B Dense model, released April 2, is posting benchmark scores that beat models with 20x more parameters: 89.2% on AIME 2026 math (vs Llama 4's 88.3%), 80.0% on LiveCodeBench v6 coding (vs 77.1%), 84.3% on GPQA Diamond science (vs 82.3%), and 86.4% on τ2-bench agentic retail (vs 85.5%). The model ranks #3 on the Arena AI text leaderboard among all open models, and with 400M+ cumulative Gemma downloads, represents a major inflection point for Google's open model strategy against Llama and GLM.