How to Run DeepSeek R1 Locally — Step-by-Step (2026)

DeepSeek R1 is a reasoning model: it “thinks” through a problem in visible steps before answering, which makes it strong at math, logic, and coding. You can run it locally for free, but the single most important thing to get right is which version you download. Almost everything labeled “DeepSeek R1” that runs on a normal computer is a distilled model — a smaller Qwen or Llama model fine-tuned on R1’s reasoning traces. The full 671B model is a different beast entirely.

This guide picks the right size for your hardware, runs it in one command, and shows how to read the output. If you have not installed Ollama yet, start with our Ollama install guide.

Pick the Right Size for Your VRAM

Every distilled tag and its download size, from the official Ollama library:

Command	Download size	Base model	Comfortable VRAM	Who it’s for
ollama run deepseek-r1:1.5b	1.1GB	Qwen	4GB+ / CPU-friendly	Tiny machines, quick tests
ollama run deepseek-r1:7b	4.7GB	Qwen	8GB	Entry-level GPUs, laptops
ollama run deepseek-r1:8b	5.2GB	Qwen3	8GB	Default (latest), best small all-rounder
ollama run deepseek-r1:14b	9.0GB	Qwen	12–16GB	The quality sweet spot for most builds
ollama run deepseek-r1:32b	20GB	Qwen	24GB	Strongest you can run on one 24GB card
ollama run deepseek-r1:70b	43GB	Llama	48GB (dual 24GB)	Multi-GPU or 48GB pro cards
ollama run deepseek-r1:671b	404GB	Full R1	Server / many GPUs	Not a consumer workload

All distilled tags ship with a 128K context window and default to 4-bit (Q4_K_M) quantization, which is the right starting point. Download size is roughly the on-disk weight; add a couple of GB of overhead plus context for the real VRAM figure. For the full memory math, see the VRAM requirements guide or run the numbers in the VRAM calculator.

The honest mapping most people need:

• 8GB GPU (RTX 4060, Arc B580 has 12GB): run deepseek-r1:8b. Good reasoning, fast.
• 12–16GB GPU (RTX 5060 Ti 16GB, RTX 4060 Ti 16GB): run deepseek-r1:14b. The best quality-for-money point.
• 24GB GPU (RTX 3090, 4090, 5090): run deepseek-r1:32b. Noticeably better at hard reasoning.

Step-By-Step Setup

1. Confirm Ollama works

ollama --version

If that errors, install Ollama first using our install guide.

2. Pull and run your chosen size

Pick the tag from the table above. For a 16GB card:

ollama run deepseek-r1:14b

The first run downloads the model, then drops you into an interactive prompt. Ask it something that rewards reasoning — a multi-step math problem or a small coding task — rather than a simple lookup, to see what the model is for.

3. Read the thinking output

DeepSeek R1 wraps its reasoning in <think> … </think> tags before the final answer. That is normal and expected: the visible chain-of-thought is the feature. If you only want the final answer, you can strip the think block in your own application, but leave it on while you are evaluating quality — it is how you tell whether the model actually reasoned or guessed.

4. Verify it loaded on the GPU

In a second terminal:

ollama ps

If the model shows as running on GPU, you are getting full speed. If it shows CPU (or a CPU/GPU split), the model did not fully fit in VRAM — drop to the next smaller size.

The 671B Reality Check

It is worth being blunt here because a lot of content blurs this line: the model that scored at the top of the benchmarks is the 671B mixture-of-experts model, and you almost certainly cannot run it at home. It is a 404GB download and needs server-class, multi-GPU memory. The distilled 1.5B–70B versions are excellent for local use, but they are smaller models taught to imitate R1’s reasoning style — not the same model at a lower quantization. Judge them on their own merits, not on the 671B’s headline scores.

Common Mistakes

• Downloading deepseek-r1:671b “to see what happens.” It is 404GB and will not run on a consumer GPU. Stick to the distilled sizes.
• Picking a size that does not fit your VRAM. A 32B model on a 16GB card spills into system RAM and crawls. Match the tag to the table above, then confirm with ollama ps.
• Judging the distill by the 671B’s benchmarks. Set expectations to the size you actually run.
• Hiding the <think> block too early. While evaluating, keep it visible — it is how you verify the model is reasoning, not pattern-matching.
• Using a reasoning model for everything. R1’s step-by-step thinking is slower and overkill for simple chat or formatting tasks. Keep a general model like Llama 3.3 around for those.

What To Do Next

• Compare against a general model — how to install Ollama covers running multiple models side by side.
• If 14B feels tight, the upgrade target is a 24GB card so 32B becomes comfortable. See the used RTX 3090 buying guide for the cheapest path to 24GB.
• Understand why the 4-bit default works so well: the quantization guide.
• Map any model to a memory target with the VRAM calculator.

Frequently Asked Questions

Which DeepSeek R1 size should I run?

Match it to VRAM: deepseek-r1:8b for an 8GB GPU, deepseek-r1:14b for 12–16GB, and deepseek-r1:32b for a 24GB card. The 14B distill is the best balance of quality and hardware cost for most people.

Is the local DeepSeek R1 the same as the one online?

No. The versions that run locally are distilled models (smaller Qwen or Llama models fine-tuned on R1’s reasoning). The full 671B model behind the headline benchmarks is a 404GB server workload.

How much VRAM do I need for DeepSeek R1 32B?

Plan for a 24GB GPU. The 32B distill is about a 20GB download at 4-bit and needs room for overhead and context on top, which is why it is the largest size that fits comfortably on a single 24GB card.

Why does DeepSeek R1 show its thinking?

That visible chain-of-thought (inside <think> tags) is the model reasoning before it answers. It is the core feature of a reasoning model and is what makes R1 strong on math and logic.

Can I run DeepSeek R1 without a GPU?

The 1.5B and 7B–8B distills will run on CPU with 8–16GB of RAM, just slowly. For anything 14B or larger, a GPU is effectively required for a usable experience.

---

Last updated: May 2026. Model tags, download sizes, base models, and context lengths reference the official Ollama DeepSeek-R1 library at time of publication. VRAM targets reference our VRAM requirements guide.