Gemma 4 code generation peaks at high temperatures

Gemma 4's high-temperature performance is a fascinating anomaly in a field where low temperature is the de facto standard for code. Higher temperatures between 1.2 and 1.5 appear to unlock better structural intuition, suggesting the model's internal reasoning is robust enough to handle high stochasticity. This behavior is likely supported by its "Thinking Mode," where internal chain-of-thought tokens provide a stable logical anchor despite "creative" sampling. With a LiveCodeBench score of 80% and a 2150 Codeforces Elo, Gemma 4 is already a powerhouse, and high-temp sampling might be the key to reaching its full reasoning potential. These findings imply that "greedy" decoding might be leaving performance on the table for frontier models with strong internal reasoning chains, suggesting a shift toward more exploratory sampling even in rigid technical domains like software engineering.