Get DeepSeek-R1 Running
After finishing this section, you will chat in the terminal with a DeepSeek running on your own computer and confirm it is using the GPU.
Before You Start
- Ollama is installed,
ollama -vshows a version number - At least 6 GB of free disk space: the deepseek-r1:8b files take up 5.2 GB
Step 1: Download and Start Chatting
ollama run deepseek-r1:8bThe first run downloads the model first. On our network, the 5.2 GB download took over 4 minutes. Once the download finishes, you go straight into the chat: type a question and press Enter; type /bye to exit.
If you just want to ask one question and leave, append it to the command:
ollama run deepseek-r1:8b "用一句话解释什么是本地部署大模型。"You should see: a thinking process first, ending with ...done thinking., followed by the answer. The answer we got in our test was:
Running large AI models directly on local devices to achieve full control and ensure data privacy.The model's answers vary from run to run; as long as the meaning is right, that's fine.
- I received an answer
Step 2: Understand the "Thinking Process"
DeepSeek-R1 is a model that thinks before answering. According to the official Ollama documentation, this kind of model outputs the reasoning process separately from the final answer. The long block you see in the command line is its thinking; only what comes after ...done thinking. is the answer for you.
When calling via the API, the thinking process is placed in a separate thinking field and the answer in the content field, as you will see in the Calling via the API section.
- I can tell the thinking process apart from the final answer
Step 3: Confirm the Model Is Installed and Using the GPU
ollama listYou should see: the downloaded models. On our machine:
NAME ID SIZE MODIFIED
deepseek-r1:8b 6995872bfe4c 5.2 GB 5 minutes agoWhile the model is running, check where it runs:
ollama psNAME ID SIZE PROCESSOR CONTEXT RUNNER UNTIL
deepseek-r1:8b 6995872bfe4c 5.6 GB 100% GPU 4096 llamacpp 4 minutes from nowThe PROCESSOR column matters most. The official explanation: 100% GPU means it is fully loaded onto the GPU; 100% CPU means it is entirely in memory; something like 48%/52% CPU/GPU means part is in memory and part on the GPU.
Note that SIZE is 5.6 GB, slightly more than the file's 5.2 GB: loading onto the GPU takes up some extra space.
Quiz timeollama ps shows PROCESSOR as 30%/70% CPU/GPU. What does this mean?
- In
ollama psmy model shows100% GPU, or I know why it doesn't
How Fast Is It
On our RTX 4080 SUPER, deepseek-r1:8b generated 172 tokens in 1.41 seconds, about 120 tokens per second. Your speed depends on your GPU; it slows down noticeably when the model doesn't fit in VRAM.
Checkpoints
- I have chatted with the local DeepSeek
- I can tell the thinking process apart from the answer
- I can use
ollama listto see which models are installed andollama psto see where they run
Didn't Work?
The download was interrupted. Run the same command again to retry the download.
PROCESSOR is not 100% GPU. Your VRAM can't fit this size; switch to a smaller one, see Choosing the Right Model Size.