What's New

AWS has made three Gemma 4 instruction-tuned models from Google DeepMind available in Amazon SageMaker JumpStart: Gemma 4 E4B, Gemma 4 26B-A4B, and Gemma 4 31B. All three models are multimodal, supporting text, image, and video inputs, with built-in reasoning, native function calling, and multilingual capabilities spanning 140+ languages. The Gemma 4 E4B variant additionally supports audio input for automatic speech recognition (ASR) and speech-to-translated-text translation, making it the most capable of the three for voice-driven workflows.

How It Works

• The Gemma 4 models are deployed through SageMaker JumpStart, which provides a managed model hub allowing customers to deploy foundation models to dedicated SageMaker endpoints with minimal configuration — either via a few clicks in SageMaker Studio or programmatically through the SageMaker Python SDK.
• Once deployed, the models expose an inference endpoint that accepts interleaved multimodal inputs (text and images in arbitrary order within a single prompt), video frame sequences, and optionally audio (E4B only).
• The built-in "thinking" mode enables chain-of-thought reasoning by having the model generate intermediate reasoning steps before producing a final answer, which can be toggled as needed.
• Native function calling support allows the model to emit structured tool-use requests, enabling integration into agentic pipelines where the model can invoke external APIs or services.
• The underlying model architecture uses a mixture-of-experts (MoE) design for the 26B-A4B variant (26 billion total parameters, 4 billion active per forward pass), balancing capability with inference efficiency, while E4B and 31B represent a compact and a full-dense model respectively.

Why It's Important

• The availability of Gemma 4 models in SageMaker JumpStart significantly lowers the barrier to deploying state-of-the-art multimodal models on AWS-managed infrastructure, eliminating the need for customers to manually source, containerize, and optimize model weights.
• The combination of image understanding (OCR, document parsing, chart comprehension, UI analysis), video analysis, audio processing (E4B), and native function calling in a single model family enables enterprises to consolidate complex, multi-step AI pipelines into fewer components.
• The 140+ language pre-training and 35+ language out-of-the-box support makes these models particularly valuable for global enterprises building multilingual applications without requiring separate localization models.
• The configurable reasoning mode also makes these models suitable for high-stakes tasks — such as legal document analysis, financial reasoning, or code review — where explainability and step-by-step logic are critical.

How It's Different

• Prior to this announcement, SageMaker JumpStart's multimodal model offerings were more limited in scope, with most available models being either text-only or supporting only basic image-text combinations.
• Gemma 4 introduces a more comprehensive multimodal stack — including video frame analysis, interleaved inputs, and audio (E4B) — within a single model family, which was previously difficult to achieve without combining multiple specialized models or using third-party APIs.
• Compared to other models available in JumpStart or Amazon Bedrock (such as Llama 3 variants or Titan models), Gemma 4's native function calling and built-in reasoning mode are first-class features rather than prompt-engineered workarounds.
• The MoE architecture of the 26B-A4B variant also differentiates it by offering a favorable capability-to-compute ratio compared to equivalently sized dense models, potentially reducing inference costs for high-throughput workloads.

When to Prefer It

• Choose Gemma 4 E4B when your application requires audio input processing, such as call center transcription, voice-driven assistants, or multilingual speech translation, especially where a compact, cost-efficient model is preferred.
• Opt for Gemma 4 26B-A4B when you need strong multimodal and reasoning capabilities at lower inference cost due to its MoE architecture activating only 4B parameters per forward pass — ideal for high-throughput production workloads with budget constraints.
• Select Gemma 4 31B for the most demanding tasks requiring maximum model capacity, such as complex document understanding, multi-step agentic workflows with function calling, or sophisticated code generation and correction.
• Across all three variants, prefer Gemma 4 over text-only alternatives when your use case involves mixed media inputs (documents, screenshots, charts, video), requires structured tool use in agentic pipelines, or demands multilingual support at scale without fine-tuning.

Availability

• The Gemma 4 E4B, Gemma 4 26B-A4B, and Gemma 4 31B models are generally available (GA) in Amazon SageMaker JumpStart as of April 29, 2026.
• Customers can access them through SageMaker Studio's Models section or via the SageMaker Python SDK.
• Specific supported AWS regions are not detailed in the announcement; availability is typically tied to the regions where SageMaker JumpStart is supported, though customers should consult the SageMaker JumpStart documentation and regional service availability pages to confirm deployment options in their target region.
• As with all JumpStart-deployed models, usage is subject to the underlying instance availability and quotas in the customer's AWS account, and costs are incurred based on the SageMaker endpoint instance type selected for deployment.
• No preview limitations or feature restrictions beyond the noted audio capability being exclusive to E4B are mentioned in the announcement.