Amazon Bedrock now carries Google DeepMind's Gemma 4 model family. Three versions are available, each aimed at a different cost and speed profile, all accessible through AWS without managing any servers.
Gemma 4 is Google's most capable open-weight model family to date. "Open-weight" means the underlying model files are publicly available, so organisations can inspect, fine-tune, and run them independently rather than relying entirely on a provider's closed system. Gemma 4 was released in April 2026 under the Apache 2.0 licence, which permits full commercial use and fine-tuning without licensing fees.
The three models on Bedrock span a wide range. The 31B is the largest and strongest, suited for complex reasoning, long documents, and code. The 26B-A4B uses a design where only a small fraction of the model activates per request, so it delivers broad knowledge at a cost closer to a much smaller model. The E2B is the smallest and fastest, built for high-volume tasks where speed and low cost matter most. All three can read both text and images, handle over 35 languages, and have a built-in reasoning mode that lets them work through complex problems step by step before giving a final answer.
The key business argument for this setup is data control. When a company sends a query to a standard cloud AI service, that data typically travels to the provider's servers and may be used to improve their models. On Bedrock, prompts and responses stay within AWS infrastructure and are not shared with Google or used for any training. This is directly relevant to businesses in sectors with strict data rules: insurance, finance, healthcare, legal, and any organisation operating under GDPR or the EU AI Act.
Regulatory pressure on AI data practices is rising sharply. Italy fined OpenAI for GDPR violations in training data handling. The EU AI Act's full enforcement for high-risk systems arrives in August 2026. Governments in Europe and Australia have started restricting which AI tools public-sector workers can use based on where data is routed. The pattern is clear: using an AI service where data routing is opaque is becoming a compliance risk, not just a preference.
Running open-weight models on a managed cloud service like Bedrock offers a middle path. Businesses get models they can inspect and, if needed, fine-tune on their own private data without building the server infrastructure themselves. AWS handles capacity, scaling, and uptime. The cost structure is pay-per-use, so there is no hardware investment sitting idle.
Bedrock now hosts models from Google, Meta, Mistral, and others, all on the same platform. That catalogue approach matters because it means a business can test different models against their own workloads and switch between them without rebuilding their setup. The API that Gemma 4 uses on Bedrock is compatible with the same interface that OpenAI's developer tools use, so teams already familiar with one can move to another with minimal effort.
For most non-technical organisations, the practical takeaway is straightforward. If your team is using any AI tool that involves sensitive customer data, internal documents, or regulated information, it is worth asking where that data goes, who can see it, and whether it is being used to train anything. Gemma 4 on Bedrock is one answer to that question: capable models, AWS-managed infrastructure, and a clear contractual guarantee that your data stays yours.