Every basic tutorial on the internet tells you to just download Ollama, type ollama run deepseek-r1, and watch the magic happen.

​What they don’t tell you is that if you don’t calculate your VRAM correctly, the model will either crash your system with a CUDA out of memory error, secretly run on your CPU at 1 token per second, or completely fill up your primary C: drive.

​Here is the actual, technical guide to deploying DeepSeek R1 locally without destroying your hardware.

Step 1: The Engine & The C-Drive Trap

Ollama is the standard engine for local AI. Go to Ollama.com, download the installer, and run it.

​The Trap: By default, Ollama installs massive model files on your primary OS drive (usually C:\Users\<user>\.ollama). If you download a 32B model, your computer will instantly run out of storage and crash.

The Fix (Windows): Before you download a model, search Windows for “Environment Variables.” Create a new System Variable named OLLAMA_MODELS and set the path to a secondary drive with massive storage (e.g., D:\OllamaModels).

​Step 2: The VRAM Math (Which Model Can You Actually Run?)

DeepSeek R1 comes in distilled sizes. You do not pick a model based on how smart you want it to be; you pick it based on your GPU’s VRAM.

• ​1.5B (Requires 4GB VRAM): The absolute minimum. Good for basic text parsing.
• ​7B/8B (Requires 8GB+ VRAM): The sweet spot for most gaming laptops (e.g., RTX 3060/4060). Excellent for coding and daily tasks.
• ​14B (Requires 16GB+ VRAM): Requires high-end hardware (e.g., RTX 4080) or an Apple M-series chip with Unified Memory.
• ​32B (Requires 24GB+ VRAM): You need an RTX 3090/4090 or a Mac Studio to run this without bottlenecking.

​Step 3: The Execution

Open your Terminal or Command Prompt. Type the command that matches your hardware and hit Enter:

ollama run deepseek-r1:7b (Replace 7b with 1.5b, 14b, or 32b based on your VRAM).

​Step 4: Advanced Troubleshooting

​Issue 1: “CUDA Out of Memory” or EOF Error

If Ollama crashes immediately after loading, your context window is eating the rest of your VRAM.

• ​The Fix: Close every other hardware-accelerated app (like Chrome or Discord). If it still crashes, you need to restrict the memory context. Run your model, and type /set parameter num_ctx 2048 to shrink the memory footprint.

​Issue 2: The Model is Agonizingly Slow (CPU Bottleneck)

Sometimes Ollama fails to detect your NVIDIA or AMD graphics card and defaults to the CPU.

• ​The Fix: Open a new terminal window while the model is generating text and type nvidia-smi. Look at the memory usage table. If your GPU memory isn’t spiking, Ollama is ignoring it. You can force GPU usage by setting another Environment Variable: CUDA_VISIBLE_DEVICES=0.

Running AI locally gives you absolute privacy and zero latency, but you have to respect the hardware constraints. If you are constantly hitting OOM errors and your current machine can’t handle the 7B models, it is time to upgrade.

Check out my guide on the Best Laptops for Local AI in 2026 to see the exact machines that can run these models flawlessly.