How to Run DeepSeek Locally: Complete Guide (2025)

1. What Is DeepSeek?

DeepSeek R1 is a reasoning-focused large language model released by DeepSeek AI in early 2025. It uses chain-of-thought reasoning — showing its work before giving a final answer — which makes it particularly strong for coding, math, and logic tasks.

Running it locally gives you full privacy (no data sent to external servers), zero API costs, and offline access. Your documents and code stay on your machine.

One important clarification upfront: the full DeepSeek R1 model (671B parameters) is designed for data centers and requires hundreds of gigabytes of VRAM. For home use, the distill versions — 8B, 14B, and 32B — are the practical choice. They retain most of R1's reasoning capability at a fraction of the hardware requirements.

2. Which DeepSeek Model Can You Run?

VRAM is the limiting factor. Here are the exact requirements for each DeepSeek variant at Q4 quantization — the standard balance of quality and size:

3. Requirements

GPU (VRAM)

System RAM and Storage

4. Step-by-Step: Run DeepSeek with Ollama

1. 1

   Check Your VRAM

   Before downloading anything, verify you have enough VRAM for your target model. On Windows, open Task Manager → Performance → GPU. On Linux/macOS with NVIDIA:

   nvidia-smi

   Use the VRAM Calculator to confirm which DeepSeek model your GPU can handle.

2. 2

   Install Ollama

   Download from ollama.com/download for macOS or Windows. On Linux, run:

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

   See the Complete Ollama Guide for detailed installation steps per OS.

3. 3

   Pull DeepSeek

   Download the model that fits your VRAM:

   8B (recommended for most setups):

   ollama pull deepseek-r1:8b

   14B (12GB+ VRAM):

   ollama pull deepseek-r1:14b

   32B (24GB+ VRAM):

   ollama pull deepseek-r1:32b
4. 4

   Run

   Start an interactive chat session:

   ollama run deepseek-r1:8b

   The <think> block is normal. DeepSeek R1 shows its chain-of-thought reasoning before giving the final answer. This can take 30–60 seconds on complex questions. It is not an error or a hang.

5. 5

   Use via API (optional, for developers)

   Ollama exposes a REST API on port 11434. Query it from any script or app:

   curl http://localhost:11434/api/generate -d '{ "model": "deepseek-r1:8b", "prompt": "Explain recursion with a simple example", "stream": false }'

5. Common Issues

6. DeepSeek R1 vs Distill: What's the Difference?

The full DeepSeek R1 (671B) uses a Mixture of Experts (MoE) architecture — it activates only a subset of parameters per token, which is why it needs less compute per inference than a dense 671B model, but still requires ~403 GB VRAM to load. That's multi-GPU server territory.

The Distill models (8B, 14B, 32B) are dense models fine-tuned from R1's outputs. They are not MoE — they are smaller, fully-dense networks trained to replicate R1's reasoning behavior using knowledge distillation. The result: they load in a fraction of the VRAM and run fast on consumer GPUs.

GPUs recomendadas para correr DeepSeek localmente

DeepSeek R1 8B Distill needs ~8 GB VRAM. 14B needs 16 GB. 32B needs 24 GB.

Precios y disponibilidad pueden cambiar. Enlaces de afiliado.