Google DeepMind's Gemma 4 models land on AWS Bedrock with MoE and reasoning

Google DeepMind has launched the Gemma 4 family of open-weight models on Amazon Bedrock, offering three instruction-tuned variants tailored to different cost and latency needs. The lineup includes the dense Gemma 4 31B (30.7B total parameters, 256K context), the mixture-of-experts Gemma 4 26B-A4B (25.2B total, 3.8B active per request, 256K context), and the compact Gemma 4 E2B (5.1B total, 2.3B effective, 128K context). All variants support built-in reasoning mode, native function calling for agentic workflows, and multimodal input across text and image, plus pretraining in 140+ languages with out-of-the-box support for 35+. Per independent benchmarks, the 31B model achieves an Artificial Analysis Intelligence Index of 39, well above the 15 median for the 4B-40B open-weights class.

Amazon Bedrock removes the typical trade-off between model access and data control. Organizations can deploy Gemma 4 as a fully managed service with inference running entirely on AWS infrastructure, backed by the platform's security and privacy controls. Prompts and completions are never used for training, and content is not shared with third parties. Users can build multimodal agents, document understanding pipelines, lightweight applications, and software engineering workflows. The open-weight nature allows independent evaluation, benchmarking, and fine-tuning on proprietary data. On-demand inference scales across three service tiers (Standard, Priority, Flex), and all variants share a common API interface including system prompts.