Gemma-4-26B-A4B with 39-69% less thinking at same accuracy

Google's Gemma-4-26B-A4B model has been optimized through ThinkingCap to reduce thinking tokens by an average of 38.9% (up to 69%) while maintaining nearly identical accuracy with only a 0.15 percentage point cost across twelve benchmarks. This fifth ThinkingCap model represents the first application of this token-reduction technique to a Google model and achieves the smallest accuracy loss of any ThinkingCap release to date, with a pro version offering even greater efficiency at up to 87.6% fewer thinking tokens.