Getting Started with DeepSeek-R1 on Tencent Cloud's High-Performance Application Service (HAI)

DeepSeek-R1 represents a significant advancement in language model capabilities, leveraging reinforcement learning to achieve impressive reasoning performance with limited labeled data. This article guides you through the process of deploying and utilizing DeepSeek-R1 on Tencent Cloud's High-Performance Application Service (HAI), enabling you to quickly test and integrate this powerful model into your applications.

What is DeepSeek-R1 and Why Use It?

DeepSeek-R1 stands out due to its enhanced reasoning capabilities achieved through extensive reinforcement learning during its post-training phase. This approach allows the model to excel in tasks such as:

• Mathematics: Solving complex mathematical problems.
• Code Generation: Writing and understanding code.
• Natural Language Inference: Understanding the nuances of human language.

Its performance in these areas rivals that of OpenAI's o1 model and other leading models.

Tencent Cloud's HAI provides a pre-configured environment for DeepSeek-R1, making it easy to get started without the complexities of setting up your own infrastructure.

Step-by-Step Guide to Using DeepSeek-R1 on HAI

Step 1: Creating a DeepSeek-R1 Application

1. Access the HAI Console: Log in to the Tencent Cloud HAI Console.
2. Create a New Instance: Click "New" to navigate to the HAI Purchase Page.
3. Configure Your Instance:
   • Application: Choose "Community Application" and select "DeepSeek-R1".
   • Region: Select a region close to your location for better network latency.
   • Compute Plan: Choose an appropriate compute plan based on the model size you intend to use. See recommended configurations below.
   • Instance Name: Customize the name, or leave it blank to use the default instance ID.
   • Purchase Quantity: The default is 1 instance.
4. Complete the Purchase: Review the configured details, select "Submit Order", and use your preferred method to complete the payment process.
5. Access Instance Details: Once the instance has deployed, click in any area of the instance to access the information detailed in the deployment overview page.

Recommended Compute Plans

For optimal performance, consider the following compute plan recommendations based on the DeepSeek-R1 model size:

• DeepSeek-R1 1.5B/7B/8B/14B: GPU Basic Plan
• DeepSeek-R1 32B: GPU Advanced Plan

Refer to the Compute Package Types documentation for detailed compute plan specifications.

Step 2: Interacting with the DeepSeek-R1 Model

Once the instance creation is complete, you’ll receive a login password via internal messaging. You now have multiple options for interacting with the DeepSeek-R1 model:

• OpenWebUI (Recommended):

  1. In the HAI console, choose "Compute Connect > OpenWebUI".
  2. In the new window, select "Start Using".
  3. Create your admin account by providing a name, email and password.
  4. Now you can begin using the model.

• ChatbotUI:

  1. In the HAI console, choose "Compute Connect > ChatbotUI".
  2. The new window will provide instructions on how to use this interface. Proceed to interact with the model.

• Terminal Connection (SSH):

  1. In the HAI console, choose "Compute Connect > Terminal Connect (SSH)".
  2. In the OrcaTerm login page, enter the login password from the account's internal messages.
  3. Once you're logged in, use the following command to load the default model:

  ollama run deepseek-r1
  

• JupyterLab:

  1. In the HAI console, shoose "Compute Connect > JupyterLab".
  2. Create a new Terminal.
  3. Run the default model with the following command:

  ollama run deepseek-r1
  

Internal Links: Explore other HAI guides such as Managing Python Virtual Environments or Building a Stable Diffusion API Service to further enhance you development.

Advanced Usage

Switching Between Different Parameter Sizes

If the default model doesn't meet your requirements, use the following commands to customize the model parameter size:

• DeepSeek-R1-Distill-1.5B: ollama run deepseek-r1:1.5b
• DeepSeek-R1-Distill-7B: ollama run deepseek-r1:7b
• DeepSeek-R1-Distill-8B: ollama run deepseek-r1:8b
• DeepSeek-R1-Distill-14B: ollama run deepseek-r1:14b
• DeepSeek-R1-Distill-32B: ollama run deepseek-r1:32b

API Calls

The environment comes pre-installed and active with Ollama serve, which can be used with with REST APIs. Refer to the Ollama API Documentation for specific instructions.

Building a Personal Knowledge Base

To build a personal knowledge base follow the steps below.

1. Cherry Studio Download: Download Cherry Studio, a multi-LLM service supporting the desktop client.
2. API Configuration: After entering the settings, choose Ollama for the model.
   • API Address: Change the default localhost to the HAI instance's public IP. Change the port from 11434 to 6399.
   • Add the deepseek-r1:7b or deepseek-r1:1.5b model ID.
3. Connectivity: Make sure you see a "Connection Successful" message when you check the API connectivity.
4. Add Local Knowledge Base Files: To use this feature, follow these steps (using the bge-m3 embed model as an example).
   • Download the Embedded Model: Click Compute Connect, then JupyterLab. Open the terminal and enter ollama pull bge-m3.
   • Add the Embedded Model: After the model download, go back to Cherry Studio. Add the bge-m3:latest model ID.

• Add/Manage Knowledge Base Files: To add files, navigate to the "Knowledge Base" page, upload your local files, and then manage your knowledge base.

Frequently Asked Questions

• What model parameter sizes are supported?

  HAI currently supports 1.5B, 7B, 8B, 14B, and 32B versions of DeepSeek-R1. 70B and 671B versions are coming soon.

• What are the port numbers for Ollama/API?

  The API port for calling Ollama in HAI is 6399. OpenWebUI uses port 6699, and ChatbotUI uses 6889. For details, please see common ports.

• How do I use the model through the API?

  Ollama serve is pre-installed and started in the instance environment. This service supports calls through the REST API. Please refer to the Ollama API documentation for specific call methods.

• What if the Ollama download speed is slow in mainland China?

  Resources in Beijing, Shanghai, and Guangzhou can be accelerated by clicking on the "Acceleration Settings" in the HAI console and enabling academic acceleration. For relevant capabilities please see enable academic acceleration.

• What should I do if I get a resource shortage message?

  Due to the popularity of DeepSeek, some regions may be out of stock, and instances cannot be created successfully. Payments will be refunded. Try changing regions or check back later.

Community and Support

For any usage issues, join the Tencent Cloud DeepSeek Deployment Exchange Group. Your suggestions and feedback are welcome!