Amazon Bedrock now hosts Hugging Face’s 83 open models, including Gemma 2

Amazon just tore down a big wall between its walled garden and the open-source AI world. As of today, 83 popular open models from Hugging Face are available directly inside Amazon Bedrock through a new Bedrock Marketplace. The move lets AWS customers deploy and run models like Google’s Gemma 2 27B Instruct using Bedrock’s existing, high-level APIs for Agents, Knowledge Bases, Guardrails, and Model Evaluations. Under the hood, the endpoints are still managed by SageMaker JumpStart, but the new integration brings those models into Bedrock’s fully managed infrastructure fold.

This is a practical power move. Previously, using an open model with Bedrock’s orchestration tools was a DIY headache involving separate SageMaker deployments and custom wiring. Now, the path is unified. You deploy from the Bedrock Model Catalog or SageMaker JumpStart, grab the endpoint ARN, and hit it with the same bedrock_runtime.converse() call you’d use for proprietary models like Claude or Titan. Pricing is straightforward: you pay standard SageMaker compute rates plus the usual Bedrock API charges.

The blog post walks through deploying Gemma 2 27B Instruct on a hefty ml.g5.48xlarge instance. The deployment process involves accepting the model’s EULA and waiting a few minutes for the endpoint to spin up. The example Python code then shows a working Converse API call, with the model correctly identifying Amazon Bedrock and Titan when asked about the company’s generative AI efforts. It’s a smooth demo, though anyone who has managed SageMaker endpoints knows the real world introduces questions about scaling, cold starts, and managing those 48xlarge bills.

What’s genuinely new here is the depth of API compatibility. By making open models play nice with Bedrock Guardrails and Agents, AWS is betting that enterprise developers want open-weight model flexibility without sacrificing the governance and tooling Bedrock provides. It’s a direct challenge to the narrative that choosing open models means building your own evaluation and safety infrastructure from scratch. The catalog’s 83 models, filterable by “Hugging Face” provider, cover a lot of ground. The cleanup instruction at the end—“Don’t forget to delete your endpoint”—is classic AWS documentation: the power is yours, along with the meter.