Gemma 4 26B (A4B)

Gemma 2 26B (A4B) delivers class-leading efficiency using a distilled architecture to outperform models twice its size.

Built on the same architecture as Gemini, the Gemma 2 26B (A4B) model utilizes a 26-billion parameter framework to punch well above its weight class. It leverages a sliding window attention mechanism and logit soft-capping to maintain high-speed inference while rivaling the performance of 70B-class models. This specific A4B iteration focuses on balancing raw throughput with sophisticated reasoning capabilities: making it the ideal choice for developers deploying on single-GPU setups like the NVIDIA A100 or H100. By prioritizing distillation techniques, Google ensures this model provides enterprise-grade accuracy for coding, mathematics, and creative agency without the overhead of massive infrastructure.

https://blog.google/technology/developers/google-gemma-2-announcement/