Run DeepSeek locally

1. Install Ollama
2. Pull and run DeepSeek-R1
3. Pick a size that fits
4. Query it over HTTP
5. Run it for a team

DeepSeek releases its models as open weights, which means you can download DeepSeek-R1 and run it entirely on your own hardware. No account, no API key, no per-token bill, and your prompts never leave the machine. The easiest way to do it is with Ollama, which handles the download and serves the model over HTTP.

#Install Ollama

If you don't already have it, install Ollama first:

curl -fsSL https://ollama.com/install.sh | sh

On macOS, brew install ollama or the desktop app work too. See run a local LLM with Ollama for the full setup.

#Pull and run DeepSeek-R1

Run the model and Ollama downloads the weights on first use:

ollama run deepseek-r1

That default tag is a distilled model that fits most machines. DeepSeek-R1 is a reasoning model, so it thinks out loud inside <think> tags before giving its final answer. Ask it something with a few steps and you'll see the working:

>>> If a server handles 200 requests per second and each takes 40ms, how many run concurrently?

#Pick a size that fits

DeepSeek ships in several distilled sizes. Choose the largest that fits in memory, since a model has to fit in RAM or VRAM to run well:

ollama run deepseek-r1:1.5b   # laptop / CPU friendly
ollama run deepseek-r1:7b     # modest GPU
ollama run deepseek-r1:14b    # mid-range GPU
ollama run deepseek-r1:32b    # high-end GPU
ollama run deepseek-r1:70b    # 40 GB+ VRAM

The small distilled versions are fine-tuned to imitate the full R1's reasoning, so you keep much of the behaviour at a fraction of the size. Check what you've downloaded and remove sizes you're done with:

ollama list
ollama rm deepseek-r1:1.5b

#Query it over HTTP

Once it's pulled, the model is available on Ollama's API on port 11434, so any script or app can call it:

curl http://localhost:11434/api/generate -d '{
  "model": "deepseek-r1",
  "prompt": "List three causes of a slow database query.",
  "stream": false
}'

The response includes the reasoning trace as well as the answer. If you only want the final answer, strip everything up to and including the closing </think> tag before you display it.

#Run it for a team

For shared use, put DeepSeek on a server with a GPU rather than a laptop. The steps are the same as any model: bind Ollama to all interfaces, then protect it with a reverse proxy and a token. The full hardening walkthrough — sizing the box, Nginx in front, TLS, firewalling the port — is in self-host an LLM on your own server.