How to Run DeepSeek Locally with Ollama

Q: How do I run DeepSeek R1 locally with Ollama?

Install Ollama, then run ollama run deepseek-r1 for the default 8B distill (~5.2 GB). Use :14b or :32b when you have more memory. Confirm the tag on the Ollama deepseek-r1 library page.

Q: Can I run DeepSeek V3 locally with Ollama?

Technically yes via ollama run deepseek-v3, but the published tag is about 404 GB with a 160K context window on the deepseek-v3 library page. For desktops and most workstations, use an R1 distill instead.

Q: DeepSeek R1 8B or 32B for local coding?

Start with 8B to validate your setup, then move to 32B if you have ~24 GB class memory and want the stronger distill tier on DeepSeek's own tables. Benchmark on your repositories rather than trusting a single leaderboard number.

To run DeepSeek locally with Ollama, install Ollama and start with a distilled DeepSeek-R1 tag that fits your machine — for most people that is ollama run deepseek-r1 (the default 8B, about 5.2 GB on disk) or deepseek-r1:14b / :32b on larger GPUs. The full 671B R1 and DeepSeek-V3 Ollama tags are ~404 GB downloads and belong on servers, not a laptop. Distill weights on the R1 card are MIT (with base-model licence notes for Qwen and Llama derivatives).

This guide answers which R1 size to pull, how DeepSeek thinking output behaves, how to raise Ollama context length, and when V3 or the full 671B checkpoint is the wrong local target. For a broader model shortlist, see open source LLM models to run locally. If you are choosing a runner instead of a model, compare Ollama vs llama.cpp vs LM Studio.

What you are actually downloading

DeepSeek-R1 is a 671B MoE reasoning model (about 37B active parameters, 128K context on the official card) trained with large-scale reinforcement learning. DeepSeek also published six dense distillations fine-tuned on R1 reasoning traces: 1.5B, 7B, 8B, 14B, 32B and 70B, based on Qwen2.5 / Qwen3 and Llama3 series bases. Those facts come from the DeepSeek-R1 model card.

Ollama's deepseek-r1 library page packages that family as simple tags. The latest / default tag is the 8B distill (Ollama documents it as the DeepSeek-R1-0528-Qwen3-8B refresh). The 671b tag is the full MoE checkpoint. Download sizes below are from Ollama's tags page — they are weight file sizes, not total runtime memory. Long context and thinking traces need extra VRAM or unified memory; see how much VRAM you need for an LLM.

Ollama tag	Approx. download	What it is (per Ollama / HF)	Practical fit
`deepseek-r1:1.5b`	1.1 GB	Distill-Qwen 1.5B	Trial / low-end laptop
`deepseek-r1:7b`	4.7 GB	Distill-Qwen 7B	Light reasoning on modest hardware
`deepseek-r1` / `:8b`	5.2 GB	Default distill (0528 Qwen3-8B)	First real local R1 on 8–16 GB class machines
`deepseek-r1:14b`	9.0 GB	Distill-Qwen 14B	Mid-range GPU / larger Mac
`deepseek-r1:32b`	20 GB	Distill-Qwen 32B	Strong dense distill on ~24 GB+
`deepseek-r1:70b`	43 GB	Distill-Llama 70B	Multi-GPU or big unified memory
`deepseek-r1:671b`	404 GB	Full R1 MoE	Server / multi-GPU only

Vendor-reported distill benchmarks on the Hugging Face card (for example AIME 2024 pass@1 of 72.6 for Distill-Qwen-32B) are useful for ranking sizes against each other. Treat them as DeepSeek-reported numbers and verify on your own math and coding prompts.

DeepSeek run locally: hardware requirements

Match tag size to free disk and memory, then leave headroom for context:

8 GB class (laptop GPU or tight RAM): start with 1.5b or 7b; try 8b only with a short context.
12–16 GB VRAM or 16–24 GB Mac unified memory: 8b is the usual daily driver; 14b is the next step if the 8B feels weak on hard reasoning.
24 GB GPU (RTX 3090/4090 class) or 32 GB+ Mac: 32b is the quality jump most people notice for math and coding distillations.
48 GB+ / multi-GPU: 70b or explore the full 671b path — still expect a 404 GB download for the big MoE tags.

Ollama also lists deepseek-v3 as a 404 GB / 160K context tag (671B total / 37B active MoE). That answers "how to run DeepSeek V3 locally" for most hobbyists: you usually do not, unless you have server storage and multi-GPU memory. Prefer an R1 distill for a desktop, or a cloud API when you need full V3 quality without the hardware.

If quant names like Q4_K_M appear in longer Ollama tag names (for example deepseek-r1:14b-qwen-distill-q4_K_M), GGUF quantization explained covers what those letters mean.

How to run DeepSeek R1 locally with Ollama

Install Ollama from the official site for macOS, Windows or Linux.
Pull and chat in one step:

ollama run deepseek-r1          # default 8B (~5.2 GB)
ollama run deepseek-r1:14b
ollama run deepseek-r1:32b

Or pull first, then run later:

ollama pull deepseek-r1:8b
ollama run deepseek-r1:8b

Check what is loaded and whether it sits on GPU with ollama ps (shows size, processor split and allocated context).

You do not need a DeepSeek API key for local tags. For OpenAI-compatible HTTP clients, Ollama's library page documents curl / Python / JavaScript chat examples against http://localhost:11434.

Raise context length (agents and coding)

The model cards advertise large windows (128K for most R1 distill tags; 160K for the 671B tag), but Ollama's context-length docs default much lower based on VRAM:

under 24 GiB VRAM → 4K context
24–48 GiB → 32K
48 GiB+ → 256K

For agents, web search or coding tools, Ollama recommends at least 64,000 tokens:

OLLAMA_CONTEXT_LENGTH=64000 ollama serve

Larger context uses more memory. If ollama ps shows heavy CPU offload, shrink context or pick a smaller tag before blaming the model.

Thinking output and sampling

R1-style models stream a visible reasoning block before the final answer. That is expected: long thoughts help hard math and coding, and they also burn tokens and latency. DeepSeek's usage recommendations on the R1 card:

temperature in 0.5–0.7 (0.6 recommended) to reduce endless repetition
avoid a separate system prompt; put instructions in the user message
for math, ask the model to reason step by step and box the final answer
when evaluating, average multiple runs rather than trusting a single sample

If answers feel shallow, enforce thinking by starting the assistant turn with a thinking marker as the card describes — otherwise some prompts skip the reasoning pattern.

Beyond Ollama: llama.cpp, LM Studio and servers

Ollama is the fastest path for "run DeepSeek locally Ollama" searches. Alternatives when you need more control:

llama.cpp / LM Studio: load community GGUF builds of the same distill checkpoints from Hugging Face Hub when you want explicit quant choice, Metal tweaks or a GUI. Same VRAM logic as above.
vLLM or SGLang: the R1 card shows serving Distill-Qwen-32B with tensor parallel for multi-GPU lab setups — better for shared OpenAI-compatible endpoints than for a single developer laptop.
Editor wiring: once Ollama is serving a tag, connect it in VS Code with how to use a local LLM in VS Code.

Licence caveats (read before shipping)

Per the DeepSeek-R1 card and Ollama's library readme:

DeepSeek-R1 series weights are MIT and allow commercial use and distillation.
Qwen-based distills start from Apache 2.0 Qwen2.5 / Qwen3 bases, then fine-tuned on R1 traces.
Llama 8B / 70B distills inherit Llama 3.1 / 3.3 licence terms from their bases.

That is the open-weight vs open-source nuance covered in open weight vs open source: permissive downloadable weights are not the same as a fully reproducible open training stack. Always read the licence file on the exact revision you redistribute.

Who local DeepSeek + Ollama is for

Good fit

Developers who want a private reasoning model without sending prompts to a cloud API
Machines in the 8–24 GB class that can run the 8B–32B distills comfortably
People comparing distill quality tiers before deciding whether a cloud V3 / full R1 endpoint is worth it

Weaker fit

Expecting full 671B R1 or V3 quality on a gaming laptop — the Ollama tags are ~404 GB for a reason
Ultra-short chat where a non-thinking instruct model is faster and good enough
Teams that need multimodal vision in the same local stack — use a multimodal family (for example run Qwen 3.6 locally) instead of text-only R1 distills

Pros and cons

Pros	Cons
One-command Ollama tags from 1.1 GB to 43 GB distills	Thinking traces add latency and token use
MIT R1 weights; clear distill size ladder	Full 671B / V3 tags are ~404 GB server downloads
Strong vendor math/coding numbers on larger distills	Llama-based distills carry Llama licence terms
Easy path into local agents once context is raised	Default Ollama context is often 4K on smaller GPUs

FAQ

How do I run DeepSeek R1 locally with Ollama?

Install Ollama, then run ollama run deepseek-r1 for the default 8B distill (~5.2 GB). Use :14b or :32b when you have more memory. Confirm the tag on the Ollama deepseek-r1 library page.

Can I run DeepSeek V3 locally with Ollama?

Technically yes via ollama run deepseek-v3, but the published tag is about 404 GB with a 160K context window on the deepseek-v3 library page. For desktops and most workstations, use an R1 distill instead.

What are DeepSeek R1 hardware requirements for local use?

Plan around the Ollama download size plus headroom for context: roughly 8B (~5.2 GB weights) for entry machines, 14B (~9 GB) for mid-range, 32B (~20 GB) for a strong single-GPU distill, and treat 671B (~404 GB) as server-class. Exact VRAM depends on context length and offload — see how much VRAM for an LLM.

Why does DeepSeek print a long thinking block?

R1-style models are trained to reason before answering. DeepSeek recommends temperature around 0.6 and warns that skipping the thinking pattern can hurt quality on hard tasks. Use thinking for math and coding; accept the extra latency.

DeepSeek R1 8B or 32B for local coding?

Start with 8B to validate your setup, then move to 32B if you have ~24 GB class memory and want the stronger distill tier on DeepSeek's own tables. Benchmark on your repositories rather than trusting a single leaderboard number.

Is DeepSeek-R1 open source?

The R1 weights are MIT on the official card, which is very permissive for use and modification. Distills based on Llama still inherit Llama licence conditions. Under a strict "open source AI" definition that requires training data and full training code, R1 is best described as open-weight — details in open weight vs open source.

Facts, prices and features in this guide were checked against official sources at the time of writing and can change. Always confirm details on the vendor's website.