DeepSeek-R1 on Azure with LangChain4j demo

Author: Julien Dubois

Introduction

DeepSeek-R1 has been announced on GitHub Models as well as on Azure AI Foundry, and the goal of this sample AI application is to demonstrate how to use it with LangChain4j and Java.

We concentrate here on GitHub Models as they are easier to use (you just need a GitHub token, no Azure subscription required), then Azure AI Foundry uses the same model and infrastructure

About this demo project

This is a Spring Boot project that demonstrates how to use DeepSeek-R1 on Azure with LangChain4j.

It contains the following demos:

• How to generate text
• A reasoning test
• How to use a chat conversation with memory of the context

Please note that despite being called a Chat Model, DeepSeek-R1 is better at reasoning than at doing a real conversation, so the most interesting demo is the reasoning test.

Those demos either run locally (using Ollama) or in the Azure cloud (using GitHub Models).

Configuration

There are 2 Spring Boot profiles, one to run locally, and one to run using GitHub Models.

Option 1 : Running DeepSeek-R1 locally

This configuration uses Ollama with DeepSeek-R1:14b, a light model that can run on a local machine

We're using the 14b version of DeepSeek-R1 for this demo: it normally provides a good balance between performance and quality. If you have less ressources, you can try with the 7b version, which is lighter, but the reasoning test will usually fail. With GitHub Models (see Option 2 below), you'll have the full 671b model which will be able to do much more complex reasoning.

It is enabled by using the local Spring Boot profile. One way to do this is to set spring.profiles.active=local in the src/main/resources/application.properties file.

Run deepseek-r1 locally using Ollama

Install Ollama to be able to run LLMs locally on your machine.

Start Ollama by typing:

ollama serve

Install deepseek-r1:14b from Ollama, available at https://ollama.com/library/deepseek-r1 by typing:

ollama pull deepseek-r1:14b

Option 2 : Running in the cloud with GitHub Models

This configuration uses GitHub Models with DeepSeek-R1:671b, the most advanced model which needs advanced GPU and memory resources

It is enabled by using the github Spring Boot profile. One way to do this is to set spring.profiles.active=github in the src/main/resources/application.properties file.

Go to https://github.com/marketplace?type=models and follow the documentation to create a GitHub token, that you'll use to access the models.

You will then need to configure an environment variable named GITHUB_TOKEN with the token value:

export GITHUB_TOKEN=<your_token>

Running the demos

Once the resources (local or GitHub Models) are configured, you can run the demos using the following command:

./mvnw spring-boot:run

Then you can access the base URL, where you find the Web UI: http://localhost:8080/.

The demos are available in the top menu.

Related samples

• DeepSeek on Azure - JavaScript demos