Amazon CEO’s Andy Jassy has shared valuable lessons learned from Amazon’s own experience between the past AWS Re: Invent. Jassy, which was brought out of this large -scale AI deployment, provided three important observation results that form a Amazon approach to the implementation of enterprise AI.
First, the cost of calculation is really important when it can be expanded with generated AI applications. People are very hungry for better prices. Second, it is very difficult to build a very good generated AI application. The third is the diversity of the model used when the builders are freely freedom to choose what they want to do. We will not surprise us because we continue to learn the same lessons over and over again.
As Andy emphasizes, a wide range of models provided by Amazon enables customers to select accurate functions that are optimal for their own needs. AWS will increase the selection of regular curated models by closely monitor both customer needs and technology advances, and include a promising new model with the favorite industry’s favorite. Enlarge. This continuous expansion of high -performance and differentiated model products helps customers at the forefront of AI innovation.
This led to China’s AI Startup DeepSeek. Deepseek released Deepseek-V3 in December 2024, and later released Deepseek-R1 and Deepseek-R1-Zero in 67.1 billion parameters, and on January 20, 2025, 1.5 billion to 70 billion parameters. The Deepseek-R1-Distill model has been released. It is based on the JANUS-PRO-7B model on January 27, 2025. The model has been published and is said to be cost -effective at 90-95 % of the same model. According to Deepseek, their models are outstanding their inference ability, which is achieved through innovative training technologies such as reinforcement learning.
Today, you can develop a DeepSeek-R1 model on Amazon Bedrock and Amazon Sagemaker AI. Amazon Bedrock is ideal for teams that promptly integrate prior trained basic models via API. Amazon Sagemaker AI is ideal for organizations that want advanced customization, training and deployment by accessing the basic infrastructure. In addition, you can use AWS TRAINIUM and AWS IMEDENTIA to develop a highly costly DeepSeek-R1-Distill model via the Amazon Elastic Compute Cloud (Amazon EC2). More.
Using AWS, you can use the Deepseek-R1 model to build a generated AI idea by building, experiments, and responsibility by building this powerful and expensive model in minimal infrastructure investment. can. By building an AWS service designed for security for security, you can promote generated AI innovation with confidence. It is strongly recommended to integrate the development of the Deepseek-R1 model with Amazon Bedrock Guardrails and add a protective layer of generated AI applications that can be used by both Amazon Bedrock and Amazon SageMaker AI.
1/ Amazon Bedrock Marketplace, Amazon Sagemaker Jumpstart, 3/ Amazon model today You can select the method to expand to S. Deepseek-R1-DISTILL model and Deepseek-R1-Distill model 4/ Amazon EC2 TRN1 instance.
Let’s walk various paths to start the AWS Deepseek-R1 model. Whether you want to build the first AI application or scaling existing solutions, these methods provide flexible starting points based on team expertise and requirements.
1. AmazonBedRock Marketplace Deepseek-R1 model
Amazon Bedrock Marketplace offers more than 100 popular, emerging, and specialized FMS, along with the current selection of the Amazon Bedrock industry. You can easily detect the model in a single catalog, subscribe to the model, and then deploy the model to the managed endpoint.
To access the Amazon Bedrock Marketplace Deepseek-R1 model, access the Amazon Bedrock Console and select the model catalog under the Foundation Models section. You can quickly find DeepSeek by searching or filtering by a model provider.
After checking out the details page of the model function and the implementation guidelines, you can provide the endpoint name, select the number of instances, select an instance, and develop the model directly.
You can also configure advanced options that can customize the security and infrastructure settings of the DeepSeek-R1 model, such as VPC networking, service roll permission, and encryption settings. The development of production must be confirmed by these settings and adjusted according to the security and compliance requirements of the tissue.
With Amazon Bedrock Guardrails, you can independently evaluate user input and model output. By filtering harmful content that is desirable in the generated AI application, the defined policy set can control the interaction between users and Deepseek-R1. The Amazon Bedrock Marketplace’s Deepseek-R1 model can only be used with the Bedrock’s Appladrail API to evaluate the user input and model response of the custom and third-party FM that can be used outside of Amazon Bedrock. For more information, implement safety measures that do not depend on the model on Amazon Bedrock Guardrails.
Amazon Bedrock Guardrails can also integrate with other rock tools, including Amazon Bedrock Agent and Amazon Bedrock Knowledge Base, to build a safer and safer AI application that matches responsible AI policies. For more information, see the AI page with AWS responsibilities.
If you use the Deepseek-R1 model using the Bedrock’s Invokemodel API and Playground Console, use the Deepseek chat template to get the optimal results. for example,
See this step-by-step guide on how to develop a Deepseek-R1 model in Amazon Bedrock Marketplace. For more information, see Deploy Models of Amazon Bedrock Marketplace.
2. AmazonsageMaker jump start Deepseek-R1 model
Amazon Sagemaker Jumpstart allows you to build a machine learning (ML) hub with FMS, an built -in algorithm, and a ML solution that can be deployed only by clicking several times. To deploy Deepseek-R1 in Sagemaker JumpStart, program through Sagemaker Unified Studio, Sagemaker AI Console, or SAGEMAKER Python SDK. You can find the EEK-R1 model.
The Amazon Sagemaker AI Console opens Sagemaker Unified Studio or SageMaker Studio. In the case of SageMaker Studio, select Jumpstart and search for “Deepseek-R1” on all public model pages.
You can select the model, select the default, and create an endpoint with the default settings. Once the endpoint becomes Inservice, you can infer a request by sending a request to the endpoint.
You can use the Amazon Sagemaker AI function, such as Amazon Sagemaker Pipelines, Amazon Sagemaker debugger, container, to derive model performance and ML operation control. This model is deployed under the AWS security environment and a virtual private cloud (VPC) control to help support data security.
Like the Bedrock MarketPalce, you can use the Sagemaker Jumpstart Applyguardrail API to separate the protective guard of the AI application generated from the DeepSeek-R1 model. You can now use Guardrail without calling FMS. This opens a door to integrate the standardized and thoroughly tested enterprise safe guard into the application flow, regardless of the model used.
See this step-by-step guide on how to deploy Deepseek-R1 in Amazon Sagemaker Jumpstart. For more information, see finding the Sagemaker Jumpstart model of Sagemaker Unified Studio. Alternatively, deploy the Sagemaker Jumpstart model in Sagemaker Studio.
3. Deepseek-R1-DISTILL model using imports of custom model models
The Amazon Bedrock Custom Model Import offers a function to import and use models that have been customized with existing FMS, without the need to manage the underlying infrastructure, and through a single serverless integrated API. With the Amazon Bedrock Custom Model import, you can import the Deepseek-R1-Distill Lalama model in the range of 1.5 billion to 70 billion parameters. As we emphasized in the blog posting on the distillation of the Amazon Bedrock model, the distillation process uses 671 billion parameters by training smaller and more efficient models and used as a teacher model, so that a larger Deepseek is larger. Immemes the operation and inference pattern of the -R1 model.
After saving these published models in the Amazon Simple Storage Service (Amazon S3) Bucket or Amazon Sagemaker registry, moved to Amazon’s foundation model Completely managed through ON It is imported into the serverless environment. bedrock. This serverless approach eliminates the need for infrastructure management while providing enterprise -grade security and scalability.
Use the Amazon Bedrock custom model imports to see this step on how to deploy the Deepseek-R1 model. For more information, import the customized model to Amazon Bedrock.
4. Deepseek-R1-Distill model using AWSTRAINIUM and AWS IRSENTIA
AWS Deep Learning AMIS (DLAMI) offers a customized machine image that can be used for deep learning on various Amazon EC2 instances, from small CPU -only instances to the latest high -power Multi GPU instance. You can get the best price performance by deploying the DeepSeek-R1-Distill model on the AWS TRAINUIM1 or AWS IMEDENTIA2 instance.
To start, access the Amazon EC2 console and launch the TRN1.32XLARGE EC2 instance using the neuron multi -framework DLAMI called Deep Learning Ami Neuron (Ubuntu 22.04).
Once connected to the booted EC2 instance, install a VLLM, an open source tool that provides a large language model (LLMS), and download the Deepseek-R1-Distill model from Face. You can use VLLM to deploy the model and call the model server.
For more information, see this stepping guide on how to develop a Deepseek-R1-Distill Lama model in AWS IMENTIA and Trainium.
Also, in Hugging Hugging, access to Deepseek-R1-Distill-Lama-8b or Deepseek-AI/Deepseek-R1-Distill-70B model card You can also do that. Select Deploy and select Amazon Sagemaker. Copy the sample code for the development of the DeepSeek-R1-Distill Lama model from the AWS IMEDENTIA and Trainium tab.
Since the release of Deepseek-R1, various guides have been posted on Amazon EC2 and Amazon Elastic Kubernetes (Amazon EKS). Here are some additional materials for you to check out:
What you need to know
There are some important things to know here.
Price setting -For public models such as Deepseek -R1, we will charge only the infrastructure price based on the infusion instance time of Amazon Bedrock Markeplace, Amazon Sagemaker Jumpstart, Amazon EC2. In the case of Bedrock Custom Model imports, it is charged only for model inference based on the number of copies of custom models charged in a 5 -minute window. For more information, see the Amazon Bedrock Pricing, Amazon Sagemaker AI price setting, and Amazon EC2 price setting page. Data Security -Use the security functions of Amazon Bedrock and Amazon Sagemaker to make data and applications safer and private. This means that the data is not shared with a model provider and is not used to improve the model. This applies to all models (dedicated and published), such as Amazon Bedrock and Amazon Sagemaker Deepseek-R1 models. For more information, see Amazon Sagemaker AI Amazon Bedrock Security and Privacy and Security.
It is now available
Deepseek-R1 is generally available on Amazon Bedrock Marketplace and Amazon Sagemaker Jumpstart. You can also use the Amazon Bedrock Custom Model Import and Amazon EC2 instance to use the DeepSeek-R1-Distill model using AWS Trainum and IDENTIA chips.
Try the Deepseek-R1 model on Amazon Bedrock Console, Amazon Sagemaker AI Console, and Amazon EC2 console. 。
-Channy
