DeepSeek-r1 14b Local Deployment Test: Detailed System Overview and Real-Time Performance Insights

Curious about how DeepSeek models perform locally?

This post will quickly guide you through the installation process for running DeepSeek models locally and share a performance test on a laptop for your reference.

Let’s get started.

Here’s my laptop system summary:

Screenshot of the Laptop System Summary

Installation

Download Ollama.https://ollama.com/
(Optional) After installation, if you would like to change the default Ollama Model Directory, you can go to Open Terminal / Command Prompt for making such change. Since I am Windows user, I opened Command Prompt and executed the following command:

setx OLLAMA_MODELS E:\Yourpath\YourFolder

Here, I changed the default directory to my E: drive for storage considerations. Large Language Models (LLMs) can quickly consume significant space, potentially filling up your C: drive. To avoid this, I recommend partitioning your drives thoughtfully and keeping large files off your C: drive whenever possible.

Open Task Manager and end the ollama.exe process if it’s running in the background. This step is crucial before downloading any models, as it ensures the environment variable changes we just made take effect.

(check out this thread if you encounter any issues with the change of the directory: https://github.com/ollama/ollama/issues/2551)

3. Go to https://ollama.com/search and select Deepseek-r1

Based on your laptop or desktop’s GPU and memory, select the option that matches your system configuration from the dropdown menu. Then, copy the command on the right and execute it in your Command Prompt.

Here, I chose to test out deepseek-r1 14b model and see how my laptop (4070 8GB) would perform.

Testing

Here’s a quick test I ran to assess the model’s utilization and performance:
Question asked: Which figure is larger, 9.11 or 9.9?

The model took about 62 seconds to complete its processing and output the answer. Check out the one-minute video for a real-time performance view:

Result Summary

The GPU utilization increased from 6% to around 30%, peaking at 48% during the session.

CPU utilization was between 30% and 40% during the session.

Both the temperature and CPU core voltage stayed within safe ranges.

(Note: Intel chips are known to face degradation issues with high voltages, so it’s essential to verify if your model is affected. For the i7 series, desktop chips like the Intel Core i7–14700K(F) and 13700K(F) have been reported to encounter such problems. Since my CPU is a laptop chip (i7–14700HX), voltages up to 1.5V are still considered safe.)

I had my usual laptop applications running during the test (browsers with 20+ tabs, communication apps, etc.). Given its performance, running the deepseek-r1 14b locally on the laptop is totally viable.

I see several benefits to running models locally, especially in terms of privacy, security, cost efficiency, and customization. deepseek-r1 makes LLM models much more affordable and accessible for average users.

I will be performing more tests on Deepseek models, so stay tuned. What’s your take on it?