Introducing Gemma 4 models on Amazon Bedrock
TL;DR
AWS made Gemma 4 available on Amazon Bedrock on June 15, 2026: three instruction-tuned open-weight models from Google DeepMind under Apache 2.0, served through the new bedrock-mantle endpoint. The lineup includes Gemma 4 31B, Gemma 4 26B-A4B, and Gemma 4 E2B. 31B is dense, 26B-A4B uses Mixture-of-Experts with 3.8B active parameters, and E2B targets low-latency use cases.
Nauti's Take
AWS frames this heavily as an infrastructure story, and that is where the value actually sits. Gemma 4 becomes more useful for many teams once it fits into existing stacks with IAM, short-lived keys, S3 images, tool calling, and familiar SDKs.
The PR-heavy part is the benchmark shine: an Intelligence Index helps with orientation, but it is not a substitute for your own tests. Bedrock users should try E2B for cheap multimodal classification and 26B-A4B for throughput before making 31B the default.
Briefingshow
The important shift is that open-weight models are moving into managed cloud infrastructure with IAM, API keys, service tiers, and Bedrock billing. That can speed up enterprise testing without forcing teams to host their own inference stack. The caveat: performance, cost, regional capacity, and the new Mantle path still need workload-specific validation.
Sources
