How to Run DeepSeek Locally: Which Model, What Hardware, Which App

The short version

You can run DeepSeek on your own computer, but only the small “distilled” versions are realistic at home. They’re free, work offline and keep your chats on your machine. They’re also noticeably less capable than the DeepSeek app.

What you can actually run

DeepSeek publishes its model weights openly on Hugging Face. The current full models are huge: the DeepSeek-V4.1-Flash-1017 files total more than 500 GB, and Ollama lists the full DeepSeek-R1 at 404 GB. That’s data-center hardware.

What fits on a normal computer is the DeepSeek-R1 distilled family. Ollama lists these sizes:

The “b” is billions of parameters. Bigger is smarter but needs more memory and runs slower.

Step by step

1. Check your memory. LM Studio recommends 16 GB of RAM or more; on Windows it also recommends a graphics card with at least 4 GB of dedicated memory. Apple Silicon Macs (M1 or newer) work well. LM Studio doesn’t support Intel Macs.
2. Pick an app. Ollama is quick if you’re comfortable typing one command. LM Studio gives you a chat window and a model browser.
3. Start small. Try the 7B or 8B version first. If it’s fast and you have memory to spare, try 14B.
4. Test it on your real tasks. Compare a few answers with the DeepSeek app to see if the quality is enough.

Prompts to copy

DeepSeek’s R1 model card recommends putting all instructions in your message rather than a system prompt:

You are helping me review a contract clause. Explain in plain English what it means,
list anything one-sided, and suggest one question to ask the other party.
Clause: [paste]

Please reason step by step, and put your final answer within \boxed{}.
A train leaves at 2:40 pm and arrives at 6:15 pm. How long is the trip?

Example

Setup: a MacBook Air with 16 GB of memory, Ollama, deepseek-r1:8b.

What to expect: it “thinks” first, showing its reasoning, then answers. It handles summaries, rewrites and simple logic well offline. On hard math or long documents it makes more mistakes than the app.

Mistakes to avoid

• Downloading 671B “because it’s the real one”. At 404 GB it won’t run on a home computer.
• Assuming a cloud tag is local. In Ollama, models tagged cloud (such as deepseek-v4.1-flash:cloud) run on Ollama’s servers, not yours.
• Expecting web search. A local model knows only what it learned in training.
• Ignoring the license of the base model. See the DeepSeek open-weight license.

Good to know

Use local for private text and offline work. Use the app for the strongest model, DeepThink and Search, when the data isn’t sensitive. Use the API to build DeepSeek into your own software. See use the DeepSeek API.

Related: run DeepSeek with Ollama and run DeepSeek in LM Studio.