In the ever-evolving landscape of artificial intelligence, DeepSeek has made a significant impact with its recent minor upgrade to the DeepSeek R1 model. This blog post delves into the installation and testing of the DeepSeek R1 0528 Qwen3 8B, a distilled reasoning model that showcases impressive performance despite its smaller size.
Understanding the Model
The DeepSeek R1 0528 Qwen3 8B is a distilled version of the larger DeepSeek R1 0528 model. Distilled models are created through a process known as knowledge distillation, where the capabilities of a larger, more powerful “teacher” model are transferred to a smaller “student” model. This allows the student model to maintain much of the performance of its larger counterpart while being more computationally efficient and easier to deploy.
The base DeepSeek R1 0528 model is a large-scale reasoning system that excels in math, programming, and logical reasoning tasks. The distillation process preserves and enhances the reasoning depth of the original model, packaging it into a more accessible 8 billion parameter architecture.
Installation Process
For this installation, we used an Ubuntu system equipped with an Nvidia RTX A6000 GPU featuring 48GB of VRAM.
We used a Python script to install the webUI version of the AI.
You can use Ollama to run this model, or with LM Studio, which is a GUI-based tool. You can go to Hugging Face to learn more about the installation.
Downloading the Model

To download the model, we utilized a script from the Text Generation Web UI directory. The model is freely available under an Apache 2 license, making it accessible without any cost. The download process revealed two shards of the model, which we waited to complete before proceeding.
Serving the Model
Once downloaded, the model was served using vLLM. The system automatically detected our GPU and started the model on the Web. Accessing the model through the Text Generation Web UI allowed us to interact with it via a web browser.

Testing the Model
Upon loading the model, we adjusted the temperature parameter to 6, as recommended by DeepSeek, to optimize the model’s output. We also set a custom system message to enhance the interaction experience.
Logical Reasoning and Math Tasks
We tested the model with a complex problem involving the optimization of widget production across multiple factories. The model demonstrated its ability to break down the problem into manageable components, perform multi-step reasoning, and provide a comprehensive solution that addressed the uncertainty in demand.
Coding Challenges
Next, we presented the model with a coding challenge involving identifying and fixing a subtle bug in a code snippet. The model showcased its systematic analysis capabilities, breaking down the problem and exploring various potential solutions. Despite not finding a bug (as there was none), the model provided insightful analysis and suggestions for improvement.
Real-World Problem Solving
Finally, we posed a real-world problem involving dating advice. The model handled the query with sensitivity and insight, acknowledging the user’s request while subtly pointing out the unrealistic nature of the expectations. It offered practical advice on how to approach the situation, emphasizing shared interests and confidence-building.
Conclusion
The DeepSeek R1 0528 Qwen3 8B model, despite its smaller size, retains the advanced reasoning capabilities of its larger counterpart. Through knowledge distillation, it achieves remarkable performance across a range of tasks, from logical reasoning and math to coding and real-world problem-solving. Its efficiency and effectiveness make it a compelling choice for those looking to deploy AI solutions with reduced computational overhead.
Whether you’re a developer, researcher, or enthusiast, the DeepSeek R1 0528 Qwen3 8B offers a powerful tool for tackling complex problems. As AI continues to advance, models like this one will play a crucial role in democratizing access to cutting-edge technology.
If you’re interested in trying out the DeepSeek R1 0528 Qwen3 8B, follow the steps outlined above to install and test it locally. The experience might just surprise you with its depth and versatility.