Best refurbished workstation for running local Copilot in 2026
Three small form factor workstations under 500 pounds that will happily host a private Copilot style assistant on your desk, plus what to look for before you buy.
Running a local Copilot in 2026 no longer requires a tower stuffed with GPUs. Quantised models in the 7B to 14B range sit comfortably in system RAM, use the integrated GPU for prompt processing, and finish coding suggestions in under a second on modern Intel silicon. You want plenty of RAM, an NVMe slot with headroom, and a CPU with AVX2 so llama.cpp, Ollama, and LM Studio behave.
Below are three refurbished workstations that hit that brief for under 500 pounds.
Why a mini workstation beats a laptop for local inference
Laptops throttle. Their fans are small, their thermal envelopes are tight, and their RAM is often soldered at 16GB. A tiny desktop workstation gives you the same silicon family, a proper heatsink, socketed SODIMMs up to 64GB, and two NVMe slots. It sits behind your monitor and pulls 35 to 65 watts under load. For an always on Copilot endpoint on your LAN, that is the sweet spot.
All three machines below are ex corporate lease returns, run Windows 11 or Ubuntu 24.04 without drama, and cost less than a mid range gaming keyboard when bought refurbished.
HP Z2 Mini G9 (from around 420 pounds)
The Z2 Mini G9 is the closest thing to a purpose built local inference box that HP makes. Depending on the configuration on eBay you get a 12th or 13th generation Intel Core i7, typically an i7-12700 or i7-13700. Refurbished units usually ship with 32GB of DDR5 SODIMM and a 512GB or 1TB NVMe drive, with a second M.2 slot free.
For local models, the Z2 Mini is the fastest of the three at prompt ingestion. The Iris Xe iGPU on the 13th gen chips accelerates llama.cpp via Vulkan, and DDR5 bandwidth (around 76 GB/s on dual channel 4800 MT/s kits) gives Mistral 7B and Qwen 2.5 7B roughly 18 to 24 tokens per second at Q4_K_M. Watch for units without the second SODIMM slot populated, since single channel memory halves throughput.
Dell Precision 3480 (from around 380 pounds)
The Precision 3480 is technically a mobile workstation, not a mini desktop, but refurbished stock is abundant and prices are keen. It ships with a 13th gen Intel Core i5 or i7, up to 64GB of DDR5, and a discrete Nvidia RTX A500 in higher trim.
The A500 has 4GB of GDDR6, which is not enough to hold a 7B model in VRAM at useful quantisation, so you will still be running most of the model through system RAM. Where the A500 earns its keep is CUDA. If your stack is Ollama with CUDA, or Continue.dev with a local Llama backend, the A500 handles prompt processing while the CPU runs generation. Expect 12 to 18 tokens per second on a Q4 7B model.
See the current stock and prices on our eBay store
Lenovo ThinkStation P350 Tiny (from around 340 pounds)
The P350 Tiny is the value pick. It uses an 11th generation Intel Core i7-11700 or i9-11900, so you are on DDR4 rather than DDR5, but the platform is proven, the chassis is one litre, and the thermals are excellent. Refurbished units regularly appear with 32GB DDR4-3200 and a 512GB NVMe for under 350 pounds.
The 11700 has AVX-512, which llama.cpp uses when compiled with the right flags. A Q4_K_M 7B model runs at 14 to 20 tokens per second, slower than the Z2 Mini but still faster than most people can read. Pick this if you want to run the assistant headless in a cupboard and expose it to other devices over Tailscale or your LAN.
The BAV Mini PC build (recommended)
A Birmingham AV custom mini build lets you specify exactly what you need without paying for silicon you will not use. A typical BAV local Copilot spec sits at around 480 pounds: an i7-12700T or i7-13700T, 64GB of DDR5, a 1TB NVMe drive, and twelve months warranty. The T suffix chips run at 35 watts, which keeps the fan quiet and the electricity bill modest.
The BIOS is tuned for sustained turbo, Ollama or LM Studio is installed ready to accept models, and the machine ships with a fresh Windows 11 Pro or Ubuntu 24.04 install as the buyer prefers. The result answers "explain this function" in the same time as a hosted Copilot, without the subscription and without the data leaving your desk.
What to check before you click buy
Three things kill local inference on a cheap refurb: single channel RAM, a slow NVMe drive, and a BIOS that caps sustained turbo. Ask the seller to confirm dual channel memory, ask for the NVMe read speed (anything under 2,000 MB/s sequential is a red flag), and check reviews for throttling under load. A twelve month warranty from a UK seller with visible feedback history beats a slightly cheaper price from a seller you cannot verify.
FAQ
Can I really run a Copilot style assistant on a 500 pound machine?
Yes, for code completion and chat on models up to about 14B parameters at Q4 quantisation. Larger models such as Llama 3.3 70B will fit in 64GB of RAM but generate at two to four tokens per second, which is fine for background summarisation but too slow for interactive coding.
Do I need a GPU?
Not for 7B to 14B models on modern Intel silicon. Integrated Iris Xe or UHD graphics handle prompt processing via Vulkan, and CPU handles generation via AVX2. A discrete GPU only becomes worthwhile at 20B and above, where you need at least 16GB of VRAM to see a benefit.
Which model should I install first?
Qwen 2.5 Coder 7B or DeepSeek Coder V2 Lite are the two best starting points for code completion in mid 2026. Both are available in GGUF format for llama.cpp and Ollama, both fit in under 8GB of RAM at Q4_K_M, and both handle TypeScript, Python, and Go well.
Will it be as fast as hosted Copilot?
For single line completions and short suggestions, yes. For long chain of thought answers or multi file refactors, hosted models are still ahead. The trade off is privacy, cost, and the fact that your assistant works offline on the train.
About Birmingham AV
We have sold more than 87,000 items on eBay since 2017, with 24,756 buyer feedbacks at 98.9% positive. Every machine ships with twelve months warranty and full UK support. Companies House 12383651, VAT GB 348755066, based in Bromsgrove Worcestershire. One of the highest volume refurbished PC operations on eBay UK, and we build every custom mini PC in house.