Gemma 4 Lands on Amazon Bedrock
Google DeepMind's open Gemma 4 family is now available on Amazon Bedrock in three variants, including a 256K-context dense model and a mixture-of-experts variant. What it is and when to reach for each one.
If you build on AWS, the practical question with any new model is rarely “is it good,” but “can I call it from the platform I already use, and which size fits my workload.” As of the AWS Weekly Roundup on June 15, 2026, Google DeepMind’s open Gemma 4 family is available on /tools/amazon-bedrock/, so you can reach it through the same managed API, security, and billing as every other Bedrock model, with no servers to run.
What happened
AWS announced that the Gemma 4 family is now available on Amazon Bedrock in three variants, each aimed at a different point on the cost, latency, and capability curve:
- Gemma 4 31B is a dense model with a 256K-token context window, suited to reasoning and coding workloads where you need to hold a lot of context at once.
- Gemma 4 26B-A4B uses a mixture-of-experts architecture, which activates only part of the network per token. That targets cost- and latency-sensitive workloads: you get a large model’s quality while paying closer to a smaller model’s compute.
- Gemma 4 E2B is the smallest variant, designed for low-latency, interactive use cases.
Gemma is Google DeepMind’s family of open models, the open-weight cousin of the proprietary Gemini line. Putting it on Bedrock means you can use an open model without standing up your own inference infrastructure.
Why it matters for builders
Two things are useful here. First, open models on a managed platform give you a middle path: more transparency and portability than a fully closed API, but without the operational burden of hosting the weights yourself. If you later want to self-host for control or cost, an open model keeps that door open in a way a closed one does not.
Second, the three variants make the cost/latency trade-off explicit. A /glossary/mixture-of-experts/ model like the 26B-A4B is the interesting one for production: it is built to keep /glossary/inference/ cheap and fast at scale, which is exactly where token spend compounds. The 256K-context 31B is the one to reach for when the task genuinely needs to see a large document or codebase at once, not by default.
The deeper point is the one the wiki keeps making: a /glossary/foundation-models/ choice is an architecture decision, not a leaderboard pick. Match the variant to the access pattern.
What to do
- Pick the variant by workload, not by size. Reach for E2B for interactive, low-latency features, 26B-A4B when you are cost- or latency-sensitive at volume, and 31B only when you truly need the 256K context.
- Try it the same way you try any Bedrock model: through the unified API, with usage-based pricing. See /tools/amazon-bedrock/ for how access, pricing, and integration work.
- Confirm regional availability and current pricing on the official AWS pages before you commit, since both vary by capability and Region.
Further reading
- Amazon Bedrock and /glossary/foundation-models/: what a managed model platform is and how to choose a model.
- /glossary/mixture-of-experts/: why an MoE model can be both large and cheap to run.
- Building RAG systems : a common place a large-context open model earns its keep.
- AWS Bedrock documentation: supported foundation models and pricing .
- Google DeepMind: the Gemma open models .
Sources
- Amazon Web Services. “AWS Weekly Roundup: AWS FinOps Agent in preview, Gemma 4 on Bedrock, Kiro Pro Max, and more (June 15, 2026).” https://aws.amazon.com/blogs/aws/aws-weekly-roundup-aws-finops-agent-in-preview-gemma-4-on-bedrock-kiro-pro-max-and-more-june-15-2026/
- Amazon Web Services. “Amazon Bedrock: supported foundation models.” https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html
- Google DeepMind. “Gemma open models.” https://ai.google.dev/gemma