A 150-DOLLAR SECOND-HAND OFFICE PC RUNS A LOCAL LANGUAGE MODEL WELL ENOUGH TO USE
A retired Dell Optiplex and two 45-dollar used workstation cards run local language models more than twice as fast as CPU-only inference, for about 165 dollars total.
by Digital Spaceport
AIHome
Built withRaspberry Pi
- difficulty
- ●●○○○
- time
- an evening
- license
- license not specified
- repo
- repo 0 stars
●●○○○ · an evening · license not specified · 0 stars · repo
Jump to section
COMPAREE VERDICT
This is not a build guide in the usual sense: it is a short parts list plus measured tokens per second for cheap used hardware, and that is more useful. Digital Spaceport ran MiniCPM-V 8B on a Dell Optiplex 7050 Mid Tower CPU-only (4.5 tokens per second) and then with used Quadro K2200, M2000 and P2000 cards and a Tesla P4. His recommended 150-dollar-class build is the 7050 Mid Tower with two bus-powered M2000 cards (11 tokens per second each), with a single Tesla P4 (28 tokens per second) as the better option if you add a cooling shroud and fan. What this gives you is a shopping list with evidence, not aspirational spec sheets. The write-up does not walk you through installing Ollama or setting up models; that lives in his separate setup guides, so you need to be comfortable with Linux basics.
IN THE REPO
GOOD TO KNOW
- —Full parts list with model numbers and measured tokens-per-second for each configuration published on Digital Spaceport site
- —Tested four used Nvidia cards (Quadro K2200, M2000, P2000 and Tesla P4) plus a CPU-only baseline
- —No repository — this is a hardware selection guide with published measurements, not a software project
- —No licence stated; content is a published article with test results
- —Software setup (Ollama, Open WebUI, llama.cpp) is not in this article; Digital Spaceport covers it in separate setup guides
- —Commercial use unrestricted — you are buying off-the-shelf parts and choosing software
Parts to buy
3 itemsFrom our check of the build. Exact quantities and part numbers are in the creator’s BOM.
Can I build this?
Build at your own risk. Projects involve tools, electronics and sometimes mains voltage — follow the creator’s safety notes.
More builds like this
All projectsGallery
Start here
Navigation into the creator’s own docs — we don’t rewrite the guide, we route you to the source.
- 1.Read the full write-up and measurements (This has the parts list, model numbers, and tokens-per-second results for every configuration tested)
- 2.Source a Dell Optiplex 7050 Mid Tower(eBay, local refurbishers, corporate surplus. The Mid Tower has room for two bus-powered cards; check the PSU, since the stock unit limits you to low-power GPUs.)
- 3.Choose the GPU route from the measured results: two Quadro M2000 cards, or one Tesla P4 plus a cooling shroud and fan(The write-up names the winner — buy that specific model, confirm VRAM size, and verify it is low-profile or that your case takes full-height brackets.)
- 4.Install Linux, Ollama, and the model of your choice(Not scripted in the article — you will need upstream Ollama docs and basic server setup knowledge.)
KNOWN ISSUES
- Buying the wrong card variant: VRAM size and power draw matter. The article tested the Quadro M2000 4GB, K2200 4GB, P2000 5GB and Tesla P4 8GB; confirm the exact model from the article before buying.
- Assuming every Optiplex 7050 is the same — some have weaker power supplies or no PCIe x16 slot. Verify the spec sheet for the specific unit you are buying.
- Underestimating the software setup — the article does not walk you through installing Ollama, setting up models, or configuring inference. If you have never done a headless Linux install, budget time for that learning curve.
- Not running the CPU baseline first — the whole point is comparing against what you already have. Test on the CPU before buying a GPU so you know the improvement is real.
- Ignoring the model size ceiling: these cards have 4 to 8 GB of VRAM each, and the article benchmarks only one 8B model at Q4. Bigger models may not fit or may spill onto the much slower CPU, so check model requirements before you buy.
- Buying from listings without photos of the actual card: confirm the exact model and VRAM before payment.
Which card won?
The Tesla P4 was fastest at 28 tokens per second but needs a cooling shroud and fan; among the Quadros the P2000 led at 20. For the 150-dollar build he recommends two M2000 cards (11 tokens per second each), and says to skip the K2200 (7).
Can I use a different base PC?
Yes, if it has a PCIe x16 slot, adequate PSU wattage for a low-power GPU, and physical clearance. The Optiplex 7050 Mid Tower is cheap and common, not magic.
Do I need the exact same CPU?
No, but the CPU-only baseline (4.5 tokens per second) was measured on his Optiplex 7050, and he mentions running an i7-7700K in it. A different processor will give different CPU-only performance, which changes the comparison.
Is this faster than a Raspberry Pi?
The article does not compare against a Raspberry Pi. The baseline it measured is the Optiplex CPU alone at 4.5 tokens per second, so the useful comparison is against the computer you already have.
Can I run Stable Diffusion or other image models on this?
Not addressed in the article, which focuses on language model inference. VRAM is the limiting factor — check model requirements.
Community builds
No community builds yet — be the first, we feature the best ones.
Discussion1
FROM THE COMPAREE TEAM
Four cards tested against a CPU baseline, tokens per second measured for each, and the Tesla P4 came out on top while two older M2000s won on price. Which card would you bet on before reading the results?
Digital Spaceport
Digital Spaceport builds and documents low-cost self-hosted infrastructure, with a focus on practical measurements rather than theoretical performance. This project is part of a wider series on running useful AI workloads on secondhand and budget hardware.
Keep building — related projects
Same category or platform, picked by how close they are to this build.
NO CAD FILESUsed cameras and an old PC become AI home security with zero monthly fees
HomeFor intermediate builders
FULL CADYou can turn a thrift-store Billy Bass into an AI that talks back
HomeFor intermediate builders
NO CAD FILESThis AI camera feeder recognizes your pet and refuses to feed the wrong one
HomeFor intermediate builders
FULL CADA used Google speaker becomes your private AI assistant with one swapped board
AIFor intermediate builders
NO CAD FILESTurn any plushie into a talking AI friend with an ESP32 chip inside
AIFor advanced builders
FULL CADA little device you tell your dream to in the morning, and it plays the dream back as video
AIFor advanced builders
DISCLAIMER
- Comparee is not the author of the projects featured here. All rights to each project belong to its creator — every page links to the original source, and we never host creators’ files.
- Information is provided without warranty and may become outdated as projects evolve. Prices are indicative bands only — always check the creator’s parts list for current costs.
- Building and operating any project is at your own responsibility. Protective equipment, safe workshop practice and compliance with local regulations are the builder’s responsibility.
CompareeTEAM1mo agoedited
Practical notes from our verification: this is a hardware selection guide with published measurements, not a step-by-step software tutorial. The article builds on a Dell 7050 Mid Tower and recommends either two Quadro M2000 cards or a single Tesla P4, and the setup of Linux and Ollama is covered in separate guides, so bring those skills or follow them. The single biggest value is the comparison: the CPU alone managed about 4.5 tokens per second, the K2200 about 7, the M2000 about 11, the P2000 about 20 and the Tesla P4 about 28 on the same 8B model, so you can see whether a cheap card is worth buying at all. Mind the power supply in the Mid Tower, which limits which cards you can fit. There is no repository and no scripts, just honest comparative testing of cheap hardware, plus a video of the build from the same author. Correction (4 October 2026): we re-checked this page line by line against the project's own repository, documentation and videos, and fixed errors in earlier versions.