Introducing Gemma 4 models on Amazon Bedrock
TL;DR
AWS is adding Google’s Gemma 4 family to Amazon Bedrock with three instruction-tuned variants: Gemma 4 31B, Gemma 4 26B-A4B, and Gemma 4 E2B. The models are open-weight under Apache 2.0 and span dense plus mixture-of-experts designs. AWS says the 26B-A4B variant activates only 3.8B parameters per token. All variants support text and image input, built-in reasoning, and native function calling. The larger models offer context windows up to 256K tokens.
Nauti's Take
This is solid infrastructure news, not a pure model breakthrough. AWS is packaging Gemma 4 as an enterprise-friendly offer: open weights without self-hosting pain, plus familiar Bedrock controls.
The PR layer is noticeable, especially around benchmarks and security framing. The useful question is whether developers can make the MoE variant beat smaller, cheaper, already integrated models in their own stack.
Briefingshow
This gives Bedrock another open model family that is positioned less as a lab option and more as a managed production building block. For teams, the interesting part is the package: open-weight licensing, long context, tool calling, multimodal input, and AWS governance. The real test is whether cost, latency, and quality beat existing options in actual workloads.
Sources
