Quick answer: The “DeepSeek R1” most people run locally is not the real 671B model. It is a set of distilled models (1.5B, 7B, 8B, 14B, 32B, 70B) built from Qwen and Llama and fine-tuned on R1’s reasoning traces. Pick the distill that fits your VRAM, pull it with one Ollama command, keep the default Q4_K_M quant, and expect a deliberate pause while it “thinks” before answering. The full 671B MoE lands around 404GB and needs a server.
DeepSeek R1 is a reasoning model: instead of answering immediately, it writes an internal chain of thought first, then commits to a response. That makes it strong on math, logic, and multi-step code, and it is why R1 became the open-weight reasoning model to beat. The confusion is that “run R1 locally” almost always means running a much smaller stand-in. This guide clears that up and gets specific about which version to pull and how it behaves.
What DeepSeek R1 actually is, and what the distills are not
The real DeepSeek R1 is a 671-billion-parameter Mixture-of-Experts (MoE) model with roughly 37B active parameters per token. The 2026 weights people cite are the R1-0528 refresh (released May 28, 2025, MIT-licensed) with a 128K-token context window and deeper reasoning than the original January 2025 release. On the AIME 2025 math test, that update pushed DeepSeek’s score from 70.0% to 87.5%, per DeepSeek’s own report.
The catch: at Ollama’s Q4 quant the deepseek-r1:671b tag is about 404GB on disk, and loading it in higher precision runs past 700GB. That is multi-GPU-server territory, not a desktop.
So DeepSeek also shipped six distilled models. Distillation here means taking a smaller, existing base model and fine-tuning it on around 800,000 reasoning and non-reasoning examples generated by the full R1. The distills inherit R1’s chain-of-thought style, but they are fundamentally the base model underneath:
- 1.5B, 7B, 14B, 32B are Qwen2.5 distillations.
- 8B is a Llama 3.1 distillation.
- 70B is a Llama 3.3 distillation.
This matters for expectations. deepseek-r1:70b is a Llama 3.3 that learned to reason like R1. It is not a shrunken 671B. The distills are genuinely good for their size (the 32B distill was reported to beat OpenAI’s o1-mini on several benchmarks, and the 14B scored 69.7% on AIME 2024), but they are a different class of model from the flagship.
The distill sizes, mapped to your VRAM
Match the tag to the card you own. Sizes are at the Q4_K_M quant Ollama pulls by default. The 1.5B and 7B figures are Ollama’s actual download sizes; the rest follow the standard 4-bit sizing rule, so treat them as close approximations and leave headroom for context.
| Ollama tag | Base model | Approx size (Q4_K_M) | VRAM tier | Notes |
|---|---|---|---|---|
deepseek-r1:1.5b | Qwen2.5 1.5B | ~1.1 GB | Any 8GB card, runs on CPU too | Tiny, fast, shallow reasoning |
deepseek-r1:7b | Qwen2.5 7B | ~4.7 GB | 8GB | Solid entry point |
deepseek-r1:8b | Llama 3.1 8B | ~5 GB | 8GB | The latest tag (~5.2GB) |
deepseek-r1:14b | Qwen2.5 14B | ~9-10 GB | 12GB tight / 16GB comfortable | Best value for one mid card |
deepseek-r1:32b | Qwen2.5 32B | ~19-20 GB | 24GB (3090/4090) | Strongest single-card option |
deepseek-r1:70b | Llama 3.3 70B | ~40-45 GB | 48GB (two cards) | Needs a real rig |
deepseek-r1:671b | 671B MoE (real R1) | ~404 GB on disk | Multi-GPU server | The actual flagship |
For the tier that fits you, the deeper picks live in our guides for 8GB cards, 12 to 16GB cards, 24GB cards, and 48GB rigs. The 32B distill sizes almost exactly like Qwen3 32B, which is no coincidence, and the 70B lands where our VRAM for a 70B model math says it should.
Exact Ollama commands
First, install Ollama. On Linux or macOS that is one line:
curl -fsSL https://ollama.com/install.sh | shThen pull and run the size that fits. One command does both:
ollama run deepseek-r1:8bSwap the tag for your tier: deepseek-r1:1.5b, :7b, :8b, :14b, :32b, or :70b. Bare ollama run deepseek-r1 gives you the latest tag, which is the 8B (~5.2GB), a safe default on an 8GB card. To pull without starting a chat, use ollama pull deepseek-r1:14b.
R1 uses more context than a normal chat model because the thinking tokens live in the same window, so raise the context on a card that can spare it:
/set parameter num_ctx 8192Quant guidance: stay at Q4_K_M
Ollama’s default tags are Q4_K_M, roughly 4-bit, and for a reasoning model that is the right call. Reasoning models lean on generating long, coherent chains, and dropping to Q3 or lower tends to make that chain wobble: more dead-ends, more self-contradiction, more wasted thinking tokens. The quality-per-byte tradeoffs are laid out in our GGUF quantization cheat sheet.
Practical rules:
- On 8 to 16GB: keep Q4_K_M and pick a smaller tag rather than a heavier quant of a bigger one. A Q4 14B beats a crushed Q2 32B.
- On 24GB: the 32B at Q4_K_M (~19-20GB) is the sweet spot. It fits with room for context; a Q5 or Q6 does not leave space for both weights and a long thinking budget.
- Do not chase Q8. The quality gain over Q4_K_M is small and it costs you the context headroom that a reasoning model actually needs.
What the reasoning tokens do to speed and context
This is the part that surprises people. Before R1 shows you an answer, it produces a <think> block: a long internal monologue where it works the problem. The refreshed R1 averages around 23,000 thinking tokens on a hard query, roughly double the original. Two consequences follow:
- Perceived speed drops. Those thinking tokens are generated at your normal tokens-per-second rate. If your rig does 30 tok/s and the model thinks for 3,000 tokens before answering, that is a real wait before the first word of the visible reply. The model is not stuck; it is reasoning. Our guide on usable tokens per second helps you calibrate once you factor the thinking phase in.
- Context fills faster. The chain of thought consumes the same window as your prompt and the answer. A short question can still burn thousands of tokens of context, which is exactly why we bumped
num_ctxabove. If the window is too small, the model runs out of room mid-thought and the answer degrades. Our walkthrough on the Ollama context window covers raising it safely.
The upside is the whole point: on math, logic, debugging, and planning, the deliberate thinking earns better answers than a same-size non-reasoning model. For casual chat that overhead feels like a tax, which is why a plain instruct model is the better daily driver and R1 is the tool for hard problems.
Honest expectations: a 7B distill is not the 671B model
If you pull deepseek-r1:7b, you get a competent Qwen2.5 7B that reasons in R1’s style. It will show its work and handle simple multi-step problems better than a same-size model without the training. It will not match the flagship on genuinely hard reasoning, broad knowledge, or reliability. The bigger the distill, the closer it gets, and the 32B and 70B are legitimately strong, but none of them is the 671B MoE. Set expectations by the base model size, not by the “R1” in the name: 8GB gets you a capable-but-shallow reasoner, a 24GB card and the 32B distill is where it starts to feel genuinely smart, and the real flagship stays a server-and-cloud model for now.
Skip the setup, if reasoning is not the point
All of the above assumes you want a reasoning model on your own hardware. If what you actually want is an uncensored companion that just talks back, the model hunt is the wrong project. Ember ships an uncensored model already chosen and tuned, downloaded once and run on your own card, with no GGUF wrangling, no quant guesswork and no VRAM math. For a reasoning workhorse, though, pull the right R1 distill for your card and keep the default quant.