Hugging Face models arrive on Amazon Bedrock Marketplace

Running open-source models on AWS used to mean choosing your poison: either wrestle with SageMaker infrastructure yourself, or stay in Bedrock and settle for whatever hosted models AWS handed you.

Now, that boundary is disappearing. Hugging Face just announced that 83 popular open-source models are officially available in the newly launched Bedrock Marketplace, rolling out across 14 AWS regions.

SageMaker muscle with Bedrock convenience

Under the hood, these deployments are actually managed by SageMaker JumpStart. But on the surface, everything lives right inside the Bedrock Model Catalog.

Developers can filter by Hugging Face, pick an open-source model like Google Gemma 2 27B Instruct, and deploy it onto a recommended GPU instance like an ml.g5.48xlarge with just a few clicks.

Once the endpoint is ready, AWS hands you a standard endpoint ARN. From there, you call it using Bedrock standard APIs, like the Converse API in the boto3 SDK, exactly as if it were a native Bedrock model.

More importantly, these Hugging Face endpoints immediately unlock Bedrock higher-level enterprise tooling, including Agents, Knowledge Bases, Guardrails, and Model Evaluations.

Billing combines the standard SageMaker compute instance cost with regular Bedrock API fees. Because these are dedicated instances rather than serverless endpoints, users still have to clean up and delete endpoints manually to stop getting billed when testing wraps up.

Why it matters

For builders, this drastically cuts down on middleware plumbing.

Plugging an open model into enterprise guardrails, retrieval-augmented generation (RAG), and agent orchestration used to require custom wrappers or third-party orchestration frameworks. Now, AWS lets you bolt enterprise-grade Bedrock Guardrails and Knowledge Bases directly onto open weights.

For AWS, this bridges an awkward gap between SageMaker and Bedrock. Competing with Google Cloud Vertex Model Garden and Microsoft Azure AI Studio requires a massive catalog, but hosting dozens of open models as pure serverless endpoints is an expensive nightmare. By delegating compute costs to SageMaker JumpStart instances, AWS cleverly expands Bedrock library without shouldering the idle GPU risk.

We think it is a smart and practical move, though not without trade-offs. Because instances are dedicated, the barrier to entry remains relatively high for casual developers running small-scale experiments, making this primarily a win for teams with steady traffic and clear enterprise compliance needs.