//
Run DeepSeek R1 reasoning model on your hardware
DeepSeek R1 is a reasoning model that rivals GPT-4 and Claude on math and coding. The distilled versions run on consumer GPUs while maintaining impressive capabilities. This guide uses Jan - a simple desktop app with built-in Model Hub.
Jan is a free desktop app that makes running DeepSeek models simple.
# Download from:
https://jan.ai/download
# Available for Windows, macOS, and Linuxš” Jan automatically detects your GPU and optimizes settings for best performance.
Open Jan and go to the Model Hub. Search for 'DeepSeek' to see the distilled R1 versions.
In Jan Model Hub, search for:
ā¢ "DeepSeek R1 7B" - Needs 8GB VRAM
ā¢ "DeepSeek R1 14B" - Needs 12-16GB VRAM
ā¢ "DeepSeek R1 32B" - Needs 24GB VRAM (best balance)
ā¢ "DeepSeek Coder" - Optimized for codingš” The distilled versions retain most of R1's reasoning capability at a fraction of the size.
Click download on your chosen model. R1 uses chain-of-thought reasoning, so responses show its thinking process.
You: What is 127 * 849?
DeepSeek R1: <thinking>
Let me break this down...
127 Ć 849 = 127 Ć (800 + 49)
= 127 Ć 800 + 127 Ć 49
= 101,600 + 6,223
= 107,823
</thinking>
The answer is 107,823.š” R1's reasoning is visible by default - this is what makes it so good at math and logic.
ā Responses are very long
ā R1 shows its reasoning by default. This is intentional and what makes it powerful. For shorter answers, ask it to be concise or just give the final answer.
ā Model runs slowly
ā 32B models need high memory bandwidth. Check that GPU acceleration is enabled in Jan Settings > Advanced. RTX 4090 is recommended for 32B.
ā DeepSeek not in Model Hub
ā Jan updates its Model Hub regularly. Check jan.ai for the latest supported models, or try searching for 'R1' instead of 'DeepSeek'.