Gemma-4-E2B-it for is now available in Amazon SageMaker JumpStart
Google DeepMind's multimodal Gemma-4-E2B-it—with vision, audio, reasoning, and function calling—is now one click away on SageMaker JumpStart.
View original announcement →Visual Summary
graph TD
A{{Gemma-4-E2B-it on JumpStart}}:::announced
B(SageMaker JumpStart):::compute
C(SageMaker Studio):::compute
D([Multimodal Processing]):::feature
E([Built-in Reasoning]):::feature
F([Native Function Calling]):::feature
G([Code Generation]):::feature
H((Developer)):::external
I(SageMaker Python SDK):::compute
H ==>|"deploys via"| C
H -->|"programmatic access"| I
C ==>|"one-click deploy"| B
I -->|"deploys"| B
B ==>|"hosts"| A
A -->|"text/image/audio"| D
A -->|"step-by-step"| E
A -.->|"agentic workflows"| F
A -.->|"generates code"| G
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
Google DeepMind's Gemma-4-E2B-it, a multimodal instruction-tuned foundation model, is now available in Amazon SageMaker JumpStart as of July 13, 2026. The model supports text, image, and audio inputs with text output, and includes a built-in reasoning mode for step-by-step problem solving. Customers can deploy it with a few clicks via SageMaker Studio or the SageMaker Python SDK.
How It Works
- Multimodal input processing: The model accepts text, image, and audio inputs simultaneously, enabling rich cross-modal understanding within a single inference call.
- Built-in reasoning mode: A chain-of-thought reasoning mechanism allows the model to internally deliberate step-by-step before producing a final answer, improving accuracy on complex tasks.
- Image understanding capabilities: Supports object detection, document parsing, screen/UI understanding, chart comprehension, and OCR out of the box without additional fine-tuning.
- Video understanding: The model can process and reason over video content, extending its multimodal capabilities beyond static images.
- Native function calling: Supports agentic workflows by enabling the model to invoke external tools or APIs directly, facilitating autonomous multi-step task execution.
- Code capabilities: Handles code generation, completion, and correction across multiple programming languages.
- Multilingual support: Operates across dozens of languages, making it suitable for globally distributed applications.
- JumpStart deployment: Customers deploy via the SageMaker Studio Models landing page or the SageMaker Python SDK, with infrastructure provisioning handled automatically by JumpStart.
Why It's Important
- Efficient edge-and-cloud flexibility: Optimized for efficient local execution, Gemma-4-E2B-it bridges the gap between on-device and cloud deployment, giving teams architectural flexibility without sacrificing capability.
- Reduced time-to-deployment: JumpStart's one-click deployment eliminates the need to manually configure endpoints, container images, or instance types, dramatically lowering the barrier to production.
- Broad task coverage in a single model: The combination of vision, audio, reasoning, code, and agentic capabilities means teams can address diverse use cases without managing multiple specialized models.
- Agentic AI enablement: Native function calling support makes this model a strong candidate for building autonomous agents that interact with external systems, a rapidly growing enterprise need.
- Google DeepMind quality on AWS infrastructure: Customers gain access to a frontier-class model from a leading AI lab while keeping data and workloads within their existing AWS security and compliance boundaries.
How It's Different
- True multimodal input (text + image + audio): Unlike many models available in JumpStart that handle only text or text-and-image, Gemma-4-E2B-it adds audio as a native input modality.
- Integrated reasoning mode: The built-in step-by-step reasoning capability is a first-class feature of the model architecture, not a prompt-engineering workaround, distinguishing it from standard instruction-tuned models.
- Optimized for efficient execution: The "E2B" designation signals an efficiency-first design, making it more cost-effective to run at scale compared to larger parameter-count multimodal alternatives.
- Native function calling: Many open-weight models require external scaffolding for tool use; Gemma-4-E2B-it supports function calling natively, simplifying agentic pipeline construction.
- Video understanding included: Video comprehension is not commonly available in similarly sized open-weight models, giving this model a differentiated capability for media and surveillance use cases.
When to Prefer It
- Document intelligence pipelines: When your application requires parsing invoices, forms, or scanned documents with OCR and layout understanding, this model handles it natively without a separate OCR service.
- Agentic and tool-use workflows: When building AI agents that need to call APIs, query databases, or orchestrate multi-step tasks, the native function calling support simplifies the architecture significantly.
- Cost-sensitive multimodal deployments: When you need multimodal capabilities but want to minimize inference costs, the efficiency-optimized design makes it preferable to larger frontier models.
- Screen and UI automation: When building accessibility tools, RPA assistants, or UI testing agents that must interpret screenshots or application interfaces.
- Multilingual enterprise applications: When your user base spans multiple languages and you need a single model to handle diverse linguistic inputs without separate localization models.
- Reasoning-intensive tasks: When accuracy on multi-step logical, mathematical, or analytical questions is critical and you want the model to show its work before committing to an answer.
- Video content analysis: When your use case involves summarizing, searching, or extracting insights from video content at scale within the AWS ecosystem.
Availability
- GA status: Generally available as of July 13, 2026; no preview or beta restrictions mentioned.
- Access method: Available through the SageMaker Studio Models landing page (SageMakerPublicHub) and the SageMaker Python SDK.
- Regional availability: Specific supported regions are not listed in the announcement; customers should check the SageMaker JumpStart regional availability documentation for the latest region support.
- Pricing model: Standard SageMaker endpoint pricing applies based on the instance type selected for deployment; no separate model licensing fee is indicated, consistent with open-weight model access in JumpStart.
- License responsibility: As with all JumpStart third-party models, customers are responsible for reviewing and complying with the applicable Google DeepMind/Gemma model license terms before use.
- Deployment prerequisite: Requires an active AWS account with appropriate SageMaker IAM permissions; access via SageMaker Studio or SDK.
