Gemma 4 models now available on Amazon Bedrock
Three Gemma 4 variants—dense, MoE, and compact—bring built-in reasoning, 256K context, and multimodal input to Bedrock's managed infrastructure.
View original announcement →Visual Summary
graph TD
A{{Gemma 4 on Amazon Bedrock}}:::announced
B(Amazon Bedrock):::compute
C([Gemma 4 31B Dense]):::feature
D([Gemma 4 26B-A4B MoE]):::feature
E([Gemma 4 E2B Compact]):::feature
F([Native Function Calling]):::feature
G([Multimodal Input]):::feature
H((Developers)):::external
I(Bedrock AgentCore):::compute
H ==>|"API requests"| A
A ==>|"hosted on"| B
A -->|"reasoning & coding"| C
A -->|"cost-efficient"| D
A -->|"low-latency"| E
A -->|"tool use"| F
A -->|"text, image, video, audio"| G
F -.->|"agentic workflows"| I
classDef announced fill:#ff9900,stroke:#ec7211,color:#fff,font-weight:bold
classDef compute fill:#e3f2fd,stroke:#1565c0,color:#1565c0
classDef storage fill:#e8f5e9,stroke:#2e7d32,color:#2e7d32
classDef feature fill:#fff3e0,stroke:#e65100,color:#e65100
classDef external fill:#f5f5f5,stroke:#616161,color:#616161
What's New
AWS has made the Gemma 4 family of open-weight models from Google DeepMind generally available on Amazon Bedrock as of June 10, 2026. The family includes three variants—Gemma 4 31B (dense), Gemma 4 26B-A4B (mixture-of-experts), and Gemma 4 E2B (compact)—each targeting different performance, cost, and latency profiles. These models bring built-in reasoning, native function calling, multimodal input, and support for 35+ languages to Bedrock's managed infrastructure.
How It Works
- Gemma 4 31B is a 30.7B-parameter dense model with a 256K-token context window, optimized for deep reasoning and coding-heavy workloads with multimodal input (text and image).
- Gemma 4 26B-A4B is a mixture-of-experts (MoE) model with 25.2B total parameters but only 3.8B active per token, enabling cost-efficient inference while retaining a 256K-token context window and full multimodal support.
- Gemma 4 E2B is a compact model with 5.1B total parameters (2.3B effective), designed for low-latency interactive applications with a 128K-token context window.
- All three variants support native function calling, structured output, built-in reasoning, and response streaming via a new Bedrock price-performance infrastructure layer.
- Multimodal input spans text and image across all variants; the announcement additionally references video and audio support in the broader Gemma 4 family.
- Models are accessed through the standard Amazon Bedrock API, benefiting from Bedrock's security, compliance, and observability tooling without requiring infrastructure management.
Why It's Important
- Open-weight flexibility: Gemma 4 models are open-weight, giving enterprises the ability to inspect, fine-tune, and deploy without proprietary lock-in, while still leveraging Bedrock's managed infrastructure.
- Broad capability coverage in one family: A single model family now covers high-complexity reasoning, cost-sensitive production workloads, and real-time interactive use cases, reducing the need to evaluate models across multiple providers.
- MoE efficiency at scale: The 26B-A4B MoE variant delivers near-large-model quality at a fraction of the compute cost, making advanced AI more economically viable for high-throughput applications.
- Production-ready agentic workflows: Native function calling and structured output support make these models directly usable in agentic pipelines and tool-augmented applications without additional prompt engineering workarounds.
- Multimodal breadth: Support for text, image, video, and audio inputs positions Gemma 4 for next-generation enterprise applications that go beyond text-only interactions.
- Long-context reasoning: The 256K-token context window on the 31B and 26B-A4B variants enables processing of large documents, codebases, or conversation histories in a single inference call.
How It's Different
- MoE architecture option: Unlike most open-weight models on Bedrock that are dense transformers, Gemma 4 26B-A4B uses a sparse MoE design, activating only 3.8B of 25.2B parameters per token—delivering better cost-per-token economics than equivalently capable dense models.
- Built-in reasoning (no separate model needed): Reasoning is natively integrated into all three Gemma 4 variants, unlike some competing offerings that require switching to a separate "reasoning" model variant.
- Tiered context windows within one family: The family spans 128K (E2B) to 256K (31B, 26B-A4B) context windows, allowing developers to right-size context capacity alongside compute cost.
- New Bedrock price-performance infrastructure: Gemma 4 runs on a new Bedrock backend specifically engineered for price-performance, differentiating it from earlier Bedrock-hosted open models that ran on general-purpose inference infrastructure.
- 35+ language support: Broader multilingual coverage than many open-weight alternatives, making it more suitable for global enterprise deployments without additional localization fine-tuning.
- Compact variant with reasoning: Gemma 4 E2B offers built-in reasoning at only ~2.3B effective parameters, a capability typically reserved for much larger models.
When to Prefer It
- Complex reasoning and code generation: Choose Gemma 4 31B when tasks require multi-step logical reasoning, advanced code synthesis, or analysis of large codebases that benefit from a 256K-token context window.
- High-throughput, cost-sensitive production APIs: Choose Gemma 4 26B-A4B when you need near-large-model quality at lower inference cost, such as batch document processing, summarization pipelines, or high-volume classification tasks.
- Real-time interactive applications: Choose Gemma 4 E2B for chatbots, copilots, or edge-adjacent workloads where sub-second latency is critical and the task complexity fits within a smaller model's capability envelope.
- Agentic and tool-calling workflows: Prefer any Gemma 4 variant when building agents that require reliable structured output and native function calling without prompt-engineering workarounds.
- Multilingual enterprise applications: Use Gemma 4 when your application must serve users across 35+ languages without maintaining separate fine-tuned models per locale.
- Multimodal document or media understanding: Use Gemma 4 31B or 26B-A4B when inputs include images alongside text, such as invoice processing, visual QA, or mixed-media content analysis.
- Open-weight compliance requirements: Prefer Gemma 4 over proprietary closed models when organizational policy requires model weight transparency, auditability, or the option for future self-hosted deployment.
Availability
- Status: Generally Available (GA) as of June 10, 2026.
- Supported AWS Regions: US East (N. Virginia), US East (Ohio), US West (Oregon), and Europe (Frankfurt).
- Pricing model: Accessed via Amazon Bedrock's standard on-demand inference pricing (pay-per-token); specific per-token rates are available on the Amazon Bedrock pricing page.
- Context window limits: Gemma 4 31B and 26B-A4B support up to 256K tokens; Gemma 4 E2B supports up to 128K tokens.
- Multimodal scope per documentation: Current Bedrock model cards confirm text and image input support; video and audio modalities referenced in the announcement may be subject to future availability updates.
- Access: Models are available through the Amazon Bedrock console and API; refer to the Google model cards in the Bedrock User Guide for model IDs and invocation details.
- No infrastructure management required: Models run fully managed on Bedrock; customers do not need to provision or manage GPU instances.
