Introducing Gemma 4 models on Amazon Bedrock

This article announces the availability of Gemma 4 models on Amazon Bedrock, a family of open-weight foundation models from Google DeepMind with built-in reasoning and multimodal capabilities.

• Three instruction-tuned variants: Gemma 4 31B (dense), 26B-A4B (mixture-of-experts), and E2B (compact) for different cost/latency profiles
• All variants support 256K token context windows, native function calling, multimodal text/image input, and built-in reasoning mode
• Access via bedrock-mantle endpoint using OpenAI-compatible APIs with no infrastructure provisioning required
• Service tiers available: Standard (on-demand), Priority (25% better latency), and Flex (discounted pricing)
• Implicit prompt caching automatically enabled to reduce latency on repeated prefixes
• Handle scaling with exponential backoff for 503 errors and gradual traffic ramps to avoid capacity issues
• Available in four AWS Regions with per-token pricing varying by model and service tier

Gemma 4 on Amazon Bedrock enables organizations to deploy open-weight models with full data protection, regulatory compliance, and operational control without managing infrastructure.