How to run DeepSeek-R1 locally on Windows (Ollama or LM Studio)

The short version

DeepSeek-R1 is DeepSeek's open reasoning model, released under the MIT licence. You can run it offline on a Windows PC with Ollama or LM Studio. Once the model is downloaded, your questions stay on your computer. That is the main difference from using DeepSeek's own app or website, whose privacy policy says data is stored in China (see is the DeepSeek app safe?).

The catch is size. The real DeepSeek-R1 has 671 billion parameters, according to DeepSeek's model card. What most people run at home is one of the smaller distilled models that DeepSeek released alongside it.

Which DeepSeek-R1 size fits your PC?

These are the download sizes on Ollama's deepseek-r1 page, with the base model each one comes from:

• deepseek-r1:1.5b: 1.1 GB, DeepSeek-R1-Distill-Qwen-1.5B
• deepseek-r1:7b: 4.7 GB, DeepSeek-R1-Distill-Qwen-7B
• deepseek-r1:8b: 5.2 GB, DeepSeek-R1-0528-Qwen3-8B. This is also what plain deepseek-r1 (the "latest" tag) gives you.
• deepseek-r1:14b: 9.0 GB, DeepSeek-R1-Distill-Qwen-14B
• deepseek-r1:32b: 20 GB, DeepSeek-R1-Distill-Qwen-32B
• deepseek-r1:70b: 43 GB, DeepSeek-R1-Distill-Llama-70B
• deepseek-r1:671b: 404 GB, the full DeepSeek-R1

DeepSeek's model cards don't publish a RAM or VRAM table, so we won't invent one. Here is what the official sources do say. Ollama's quickstart explains that a model runs fastest when it fits in your graphics card's memory (VRAM). With less VRAM, Ollama uses system RAM, "but responses may be slower". LM Studio recommends at least 16 GB of RAM and 4 GB of dedicated VRAM on Windows.

A practical way to choose: the download size is roughly how much memory the model takes up once loaded, and the conversation (context) needs a little more on top. So pick a size that is clearly smaller than your free GPU memory, or your free RAM if you have no graphics card. On an 8 GB laptop, start with 1.5b. With 16 GB of RAM, 7b or 8b is a sensible first try. This is a rule of thumb, not a test result. Use ollama ps (below) to see what actually happens on your PC.

Method 1: run DeepSeek-R1 with Ollama

1. Install Ollama for Windows. Our step-by-step guide is here: how to install Ollama on Windows.
2. Open PowerShell or Command Prompt.
3. Run ollama run deepseek-r1:8b, or replace 8b with the size you chose. The first run downloads the model.
4. When the >>> prompt appears, type your question. R1 is a reasoning model, so it "thinks" before it answers, and the reasoning is shown first.
5. Type /bye to exit. Next time the model starts without downloading again.
6. To check how it's running, open a second window and run ollama ps. "100% GPU" is ideal. "100% CPU" means it is running from RAM and will be slower.

Ollama's page notes that DeepSeek-R1 received a minor upgrade to DeepSeek-R1-0528. To update a copy you downloaded earlier, run ollama pull deepseek-r1.

Method 2: run DeepSeek-R1 with LM Studio

1. Download LM Studio from lmstudio.ai and install it. On x64 PCs, LM Studio requires a CPU with AVX2 support.
2. Open the Discover tab (press Ctrl + 2) and search for deepseek-r1.
3. You'll see download options with names like Q4_K_M or Q8_0. These are quantizations: compressed copies of the same model. LM Studio's advice is to "choose a 4-bit option or higher if your machine is capable enough".
4. When the download finishes, go to the Chat tab, open the model loader (Ctrl + L) and pick the model.
5. Start chatting. LM Studio's docs say that once a model is on your machine, "nothing you enter into LM Studio when chatting with LLMs leaves your device".

Settings DeepSeek recommends

DeepSeek's model card for the R1 series recommends a temperature between 0.5 and 0.7, with 0.6 as the ideal, "to prevent endless repetitions or incoherent outputs". For the original R1 models it also says to avoid a system prompt and put every instruction in your own message. The newer DeepSeek-R1-0528-Qwen3-8B card says a system prompt "is supported now". In Ollama, temperature is a model parameter you can set in a Modelfile (PARAMETER temperature 0.6). For maths problems, DeepSeek suggests adding "Please reason step by step, and put your final answer within \boxed{}."

Troubleshooting

• Very slow answers: the model is too big for your VRAM and is spilling into RAM. ollama ps will show CPU in the PROCESSOR column. Use a smaller size, such as 8b instead of 14b, and close other heavy apps.
• Not enough disk space: models are stored in your user folder by default. In Ollama, set the OLLAMA_MODELS environment variable to a folder on another drive and restart Ollama. In LM Studio, change the models folder in My Models.
• It repeats itself or rambles: set the temperature to around 0.6, as DeepSeek recommends.
• Long "thinking" before every answer: that's normal for a reasoning model. Smaller distills think faster but are less capable.
• LM Studio won't start on an old PC: check that your processor supports AVX2. LM Studio lists it as required on x64.

Licence

DeepSeek's card says the R1 code and weights are under the MIT licence and allow commercial use. The distilled models also carry their base model's licence: Apache 2.0 for the Qwen-2.5 ones, and Meta's Llama 3.1 or 3.3 licence for the Llama ones.