How to Run DeepSeek Locally on Windows

If local models are new to you, our overview of running local AI models on Windows is a good primer; this page focuses on the DeepSeek family.

What DeepSeek models are

DeepSeek is a series of open-weight models from the Chinese AI lab of the same name. The line includes general-purpose chat models and a family of reasoning models trained to work through problems step by step. DeepSeek drew a lot of attention for reaching strong results with open weights, and because its models are permissively licensed for commercial use and further training.

The important thing to understand before you download anything is scale. DeepSeek's flagship models are very large mixture-of-experts (MoE) networks — hundreds of billions of parameters in total. They are designed for data-centre hardware and will not fit on a single home PC. So while you will see the full model discussed everywhere, that is not what you run on a laptop. What you run at home is one of the smaller distilled variants described next.

The distilled, smaller variants

To make its reasoning ability usable on ordinary hardware, DeepSeek released distilled models. Distillation trains a smaller model to imitate the behaviour of a much larger one. In DeepSeek's case, the reasoning behaviour of the big model was distilled into compact open models built on Qwen and Llama base architectures.

These distilled models come in a spread of sizes — commonly 1.5B, 7B, 8B, 14B, 32B and 70B parameters — with the smaller ones based on Qwen and the larger ones on Llama, a split you can see on the official DeepSeek-R1 model page in the Ollama library, which credits the 70B distill to Llama 3.3 and the smaller ones to Qwen. They keep much of the step-by-step reasoning style of the full model while being small enough to run on a desktop GPU, and in the smallest sizes even on CPU. For anyone wanting to try DeepSeek locally, a distilled build is the answer.

Picking a size vs your VRAM and RAM

As with any local model, memory decides what runs well — chiefly your GPU VRAM, with system RAM as the fallback. A quantized model needs roughly as much memory as its file size on disk, plus headroom for the conversation. Here is a rough guide for the distilled DeepSeek sizes:

If a model is bigger than your VRAM, Ollama and LM Studio can offload part of it to system RAM — it still runs, just more slowly. The sensible plan is to start with the 7B or 8B distill, confirm it feels responsive, and move up only if you need more. For the full memory picture, see our guide on how much VRAM you need for local AI.

Run DeepSeek with Ollama

Ollama is a native Windows app that runs models from a command line and serves them through a local API. If it is not installed yet, our guide to installing Ollama on Windows covers it; the short version is below.

1. Download and run the official Ollama Windows installer.
2. Open Windows Terminal or PowerShell.
3. Pull and run a distilled DeepSeek in one command — ollama run is the documented way to start a model in Ollama's CLI reference — choosing a size tag that fits your memory:

   ollama run deepseek-r1:8b

   Swap 8b for 1.5b, 7b, 14b, 32b or 70b depending on your hardware; all of those size suffixes are listed on the deepseek-r1 tags page.
4. Type a prompt and press Enter. On a reasoning model you will see a thinking section appear before the final answer (more on that below). Type /bye to leave the chat.
5. Manage downloads to reclaim disk space:

   ollama list
   ollama rm deepseek-r1:8b

Run DeepSeek with LM Studio

Prefer a graphical app? LM Studio has a model browser and chat window and needs no terminal.

1. Install LM Studio for Windows from its official site.
2. Use search to find DeepSeek. LM Studio lists builds from public model hubs and shows a fit hint for each quantization against your machine's memory.
3. Pick a distilled build that fits — a 7B or 8B at 4-bit is a good first pick — and click Download.
4. Open a chat, load the model and start typing. Everything runs locally.

Torn between the two tools? Our side-by-side comparison of Ollama vs LM Studio lays out the trade-offs; many people keep both.

Reasoning-model notes

DeepSeek's reasoning models behave a little differently from an ordinary chat model, and it helps to know what to expect. They are trained to work through a problem step by step before giving a final answer, and they expose that working as a visible thinking or reasoning section that appears first.

• Answers are longer and slower. Because the model generates its reasoning as well as the answer, each reply takes more time and produces more text. On a small local build that is noticeable.
• It shines on multi-step problems. For logic, maths, planning and anything that benefits from working-out, the step-by-step approach tends to be more reliable than a model that answers in one shot.
• It can feel verbose for quick questions. For a one-line factual query, all that thinking is overhead. If you mostly want fast, direct chat, a non-reasoning general model may suit you better — see our roundup of the best local LLMs for your PC for alternatives.
• The thinking is not a guarantee. Visible reasoning makes a model's process easier to follow, but it does not make it correct. Local models still make mistakes; check anything that matters.

Privacy: local vs the DeepSeek app

This is where running DeepSeek yourself has a clear, concrete advantage. The DeepSeek phone and web apps are a cloud service: what you type is sent to DeepSeek's servers, processed there, and handled according to its privacy policy and the jurisdiction it operates under. That is fine for casual use, but it is not private in the way many people assume, and it has prompted scrutiny from regulators and organisations over where the data goes.

When you download an open DeepSeek model and run it locally with Ollama or LM Studio, the picture changes completely. Once the file is on your disk, inference happens entirely on your machine — your prompts and the model's replies never leave the device, no account is required, and no connection is needed to chat. The only network step is the one-time download of the model itself. That makes a local distilled DeepSeek a strong fit for confidential documents, sensitive work and offline use, in a way the cloud app is not.

For the wider context on Windows and on-device AI — NPUs, Copilot+ PCs and Microsoft's own stack — see our overview of Windows AI.