Running DeepSeek R1 locally on a refurbished PC

DeepSeek R1 changed the calculus on local AI. Before it landed, running a capable reasoning model at home meant buying a £2,000 GPU or accepting painfully slow CPU output. The distilled variants of R1 sit in a much friendlier place. On a mid range refurbished tower you can get useful, private, offline reasoning at conversational speed for under £400 all in.

What the distilled variants actually are

DeepSeek R1 in its full form is a 671 billion parameter mixture of experts model. Nobody is running that at home. The variants that matter for local use are the distilled ones, where the reasoning behaviour of the full model has been transferred into smaller open weight base models (Qwen and Llama).

The four you will see on Ollama and LM Studio are:

• R1 Distill Qwen 1.5B: tiny, fast, useful for text cleanup and simple structured extraction.
• R1 Distill Qwen 7B: the sweet spot for most home users. Good reasoning, quick responses.
• R1 Distill Qwen 14B: noticeably stronger on multi step maths and code, still practical.
• R1 Distill Qwen 32B: the heavyweight. Approaches GPT-4o mini on several benchmarks.

Each is available in several quantisation levels. Q4_K_M is the sensible default: it cuts memory use to roughly half of FP16 with a barely measurable quality hit.

Hardware requirements by variant

The single most important number is available RAM (or VRAM if you are running on a GPU). Everything else is secondary. These are the practical minimums at Q4_K_M quantisation, based on real memory footprints not marketing claims:

• 1.5B: 2GB RAM minimum, 8GB comfortable. Runs on almost anything, including a 2015 laptop.
• 7B: 8GB RAM minimum, 16GB comfortable. This is where a modern refurbished tower shines.
• 14B: 16GB RAM minimum, 32GB comfortable. Needs a decent CPU or a mid range GPU to stay responsive.
• 32B: 24GB RAM minimum, 48GB comfortable. Realistically wants a GPU with 20GB+ of VRAM, or patience.

CPU inference speed scales with memory bandwidth more than clock speed. A Ryzen 7 5700G with DDR4-3200 will outperform an older i9 with DDR3 here. Add an RTX 3060 12GB or a used RTX 4060 Ti 16GB and the 14B model moves into properly interactive territory (30+ tokens per second).

The Ryzen 7 baseline: what to buy

For most people, the honest answer is a refurbished Ryzen 7 tower with 32GB of RAM and an NVMe SSD. It runs the 7B model at conversational speed on CPU alone, handles the 14B at usable speed, and leaves headroom to drop in a used GPU later if you want to push into 32B territory.

The Birmingham AV Ryzen 7 workstation sits exactly in that spec: 32GB DDR4, fast NVMe, spare PCIe slot, and PSU headroom for a mid range GPU upgrade.

View the BAV Ryzen 7 workstation on eBay

Pricing varies by configuration: the entry point covers a Ryzen 7 with 16GB and a 512GB NVMe, the upper tier bumps to 32GB RAM, 1TB NVMe, and a discrete GPU option. Pick the 32GB variant if local AI is a priority. RAM is the ceiling on which models you can load, so it is the wrong place to economise.

Practical use cases

There is a habit of talking about local LLMs as if the only benchmark is whether they can replace ChatGPT. That is the wrong frame. The interesting question is which jobs are worth doing locally that you would never send to a cloud API.

Private document analysis. Feed the model your accounts, contracts, or medical notes and ask questions. Nothing leaves the machine. R1 distilled 7B is more than adequate for extracting figures and summarising.

Code review on proprietary codebases. The 14B distill is genuinely useful for spotting bugs on a codebase you are not allowed to paste into a cloud service.

Structured data cleanup. Take a spreadsheet of messy supplier names, SKUs, or address data and normalise it. The 7B handles this at speed and never rate limits you.

Always on background tools. Because there is no API cost, you can run the model against every email, every saved article, every meeting transcript. Cost per query is electricity, not tokens.

What you actually give up

Local models are not a free lunch. The 7B distill will make mistakes GPT-4 class models do not make. Long context handling is weaker. Web search and tool use require wiring up Open WebUI or a custom agent, they do not come out of the box. If your workflow needs state of the art benchmarks, run it in the cloud. If your workflow needs privacy, cost predictability, or offline availability, local wins by a wide margin.

Getting started in ten minutes

Install Ollama (or LM Studio for a GUI), open a terminal, and run ollama pull deepseek-r1:7b. That is the whole setup. From there you can chat directly, point Open WebUI at it for a browser interface, or wire it into VS Code via Continue. The whole stack is open source and does not phone home.

FAQ

Can I run the 14B model on a Ryzen 7 without a GPU?

Yes, at around 6 to 10 tokens per second on a 5700G with fast DDR4. That is slower than reading speed but workable for background tasks. If you want conversational speed on the 14B, add a 12GB GPU.

Do I need Linux?

No. Ollama and LM Studio both run fine on Windows 11. Linux is marginally faster on CPU inference (around 5 to 10 percent) due to scheduler differences, but the gap is not worth switching operating systems over.

How much RAM do I really need?

For the 7B distill you can get away with 16GB, but 32GB is the sensible answer for a machine you want to grow with. You want headroom to run the model alongside a browser, an IDE, and whatever else you actually do.

Is a refurbished PC really the right choice for this?

For local LLM work, yes. The bottleneck is memory capacity and bandwidth, both cheap on a two or three year old business tower. You get silicon that is genuinely good enough, at a fraction of new prices, with a warranty attached.

About Birmingham AV

We have sold more than 87,000 items on eBay since 2017 and hold 24,756 buyer feedbacks at 98.9% positive. Every machine ships with a twelve month warranty. We are Birmingham AV Ltd, Companies House 12383651, VAT GB 348755066, based in Bromsgrove, Worcestershire. One of the highest volume refurbished PC operations on eBay UK.