DeepSeek R1 represents a new wave of large language models (LLMs) built for advanced reasoning tasks. These “reasoning” models use Chain-of-Thought (CoT) processing, where the model plans and reflects before generating a final answer. This makes them exceptionally good at solving complex problems—especially in math, logic, and science—even if responses take a little longer to appear.
DeepSeek has distilled R1 into smaller, optimized models that retain strong reasoning abilities. You can now run these models locally on AMD Ryzen AI processors and Radeon graphics cards using LM Studio. Watch it in action: See a real-time demo on YouTube
What Makes Reasoning Models Different?
Unlike conventional LLMs that deliver responses in a single step, reasoning models generate a chain of thought before producing a final answer. This internal process can span thousands of tokens as the model explores different angles of a problem.
In LM Studio, this chain of thought is visible through the “thinking” window. You can expand it to see how the model reasons through a task, providing transparency and insight into how it reaches conclusions.
The result is deeper, more accurate responses—especially valuable when dealing with nuanced or technical topics.
Getting Started: DeepSeek R1 on AMD Ryzen AI & Radeon GPUs
Running DeepSeek R1 locally is straightforward, provided your system meets a few requirements. Watch the setup guide: LM Studio + DeepSeek R1 Tutorial
Requirements
- AMD Adrenalin Driver 25.1.1 or newer
Download the 25.1.1 Optional Driver here - LM Studio version 0.3.8 or newer
Download from lmstudio.ai
Step-by-Step Setup
- Update Your AMD Drivers
Make sure you’re running Adrenalin version 25.1.1 or later to enable AI acceleration features. - Install LM Studio
Download LM Studio from the link above, install it, and skip the onboarding screen. - Open the Discover Tab
Launch LM Studio and click the Discover tab to view available models. - Select a DeepSeek R1 Distill Model
Start with smaller models like Qwen 1.5B for fast performance. Larger models offer deeper reasoning but require more resources. Refer to the official chart for AMD’s maximum recommended model sizes.AMD recommends using Q4_K_M quantization for all distills. - Choose Quantization and Download the Model
On the right side of the screen, select Q4_K_M and click Download. - Load the Model in the Chat Tab
Go to the Chat tab. From the dropdown, select your DeepSeek R1 model. Enable the option to manually select parameters. - Enable GPU Offloading
Move the GPU offload slider all the way to the right to maximize hardware acceleration. - Click ‘Model Load’
The model will now initialize and load directly on your AMD hardware. - Begin Interacting
You’re ready to use a powerful reasoning model locally, without relying on the cloud.
With these steps, you’re set to explore advanced reasoning capabilities powered entirely by AMD. Whether you’re tackling complex technical problems or just want a more thoughtful AI assistant, DeepSeek R1 distillations deliver strong performance and insight—right from your desktop.
