A budget is a ceiling you choose before the forum does
A local AI setup under $600 means a computer that can run a small model with the network off, bought or assembled without pretending you are building a data center. Six hundred dollars is a ceiling for people who do not already have a usable machine, not a quote and not a dare. If you already own a decent laptop, your setup cost may be zero plus an evening. Spend the money only on the wall you hit.
The hardware guide explains how to turn a model file's size into a memory budget. The beginner local guide is the first software path. This article is the buying and assembling order so the ceiling holds. Prices on shelves move. Check them the day you buy. Nothing here is a street price.
Start with the model, not the case
Write down the job. Private notes, a small coding helper, a rough image, transcription of files you will not upload. Each job has a smallest model that is worth running. That file's size, plus the system you still want to use at the same time, is the memory number. If you skip this step you will buy a graphics card shaped object because a thread was excited.
For text, a small instruct model is the honest target on a budget. For images, a small diffusion model is a sketch engine, not a studio. For transcription, a Whisper-class model on CPU is often enough and does not need a new card at all. If your job is transcription, do not buy a GPU because a chatbot guide told you to.
Use what you have, in this order
- Try the model on the computer you own. A used-feeling laptop that loads the small file is a successful setup.
- If it loads but the machine swaps and becomes unusable, the wall is memory. On a desktop you may be able to add RAM. On a laptop you usually cannot. Believe the manual.
- If it does not fit at all, you need a different computer or a smaller file. Try the smaller file before the different computer.
- Only then look at a purchase under your ceiling.
A Mac with enough unified memory is a common "already own it" answer. The Mac Llama guide is that path and it may cost nothing. A Windows desktop with a free RAM slot is the other common path. A new high-end card is not required to learn.
Where a purchase under the ceiling usually goes
If you must buy, put the money into memory and a machine that can hold it, not into RGB and not into a power supply sized for a card you do not have.
| Piece | Why it gets the money | Why it does not |
|---|---|---|
| Used business desktop or mini PC with upgradeable RAM | You can reach a comfortable memory number without a boutique case | It will not train large models. You are not training large models. |
| RAM to fill the slots the manual allows | Small models plus a browser need room | Buying RAM you cannot install |
| A modest used GPU, only if a specific model requires it and the power supply and case can take it | Some image models and some quantizations want it | Buying a card that costs the whole ceiling and starves the rest |
| A big new power supply "for later" | Later is how budgets die | You do not have the card |
| A second monitor | Nice, irrelevant to the model | It does not make answers better |
Used markets vary. Look at the seller's return window more than at the benchmark screenshot. A machine you cannot return is not a bargain if it fails the memory test in week one.
Stay under the ceiling by writing the number at the top of the note and subtracting as you go. When the remainder is too small for the next part, stop and use the smaller model. The ceiling is the point.
Software that does not add a second bill
Ollama or LM Studio for text. A local Whisper app for audio. A small local image UI if, and only if, images are the job. One of each. Three chat apps downloading three copies of the same weights will fill a drive and blow the "free software" story.
Turn off cloud fallbacks. Confirm with the network off, as the Mac guide describes. A budget local setup that secretly calls an API is a hosted plan with extra hardware.
Delete models you are not using. Disk is part of the budget. A 1 TB drive full of abandoned weights is a purchase you already made.
What this setup will not do
It will not match a frontier hosted model at open-ended research or huge context. Use a hosted tool, on purpose, for text that is allowed to leave. The assistant comparison is that choice. Local is for the text that should stay.
It will not fine-tune a serious model. Fine-tuning wants data discipline more than it wants a bargain GPU. Read the fine-tune guide before you buy anything "for training." Most people should not buy anything for training.
It will not be silent, small, and fast all at once if you insist on a large file. Pick two.
A weekend plan
Saturday morning: write the job and the smallest model name. Saturday afternoon: try it on the current computer, network off. Sunday: if it failed for memory, price one upgrade that fixes that wall and nothing else. If the upgrade does not fit under the ceiling, keep the smaller model or wait. Do not "stretch" the ceiling for a card a video recommended. The video does not know your job.
Write the final inventory in four lines: machine, memory, model file, folder. That note is the setup. The boxes are not.
Frequently asked questions
Q: Can I build a local AI PC for under $600?
A: Sometimes, if you buy used and your job fits a small model. Sometimes you already own the computer and should spend nothing. Check current prices and the model file size before you believe any build list, including this one.
Q: Do I need a GPU?
A: Not for a first small text model or for basic transcription. You need a GPU when a specific model will not run acceptably without one. Prove that on the small file first.
Q: Is a used Mac better than a used Windows desktop?
A: The better one is the one that has enough memory for your file and that you can actually buy inside the ceiling. Apple's memory is not upgradeable later. A desktop's RAM sometimes is. Read the manual for the exact machine.
Q: What if $600 is my whole computer budget, not just the AI budget?
A: Then the computer has to do your normal work too. Buy the reliable used office machine with memory, and run a small model on it. Do not sacrifice a working daily driver for a card.
Q: Can I run this on a laptop?
A: Yes, if the laptop has the memory and you accept fan noise and a smaller model. Do not buy a gaming laptop at the edge of the ceiling for a job a used desktop would do more quietly.
Q: Are cloud GPUs cheaper?
A: Renting can be cheaper for a one-time experiment and worse for private notes, because the data is on someone else's machine. This article is about local. If the text can be uploaded, you may not need the hardware at all.
Q: How much electricity will it use?
A: More when a GPU is busy, less when you run a small model in short sessions. It is not free. It is usually cheaper than a stack of subscriptions if you actually turn the machine off. The subscription guide is the other side of that comparison.
Q: What should I skip to stay under the ceiling?
A: Extra case fans you do not need, a oversized power supply, a second GPU, and any part bought "for the next model." The next model can wait until this one does a job.
The ceiling is the feature
A local setup under $600 is mostly restraint. Name the job, try the computer you have, buy one piece that fixes a measured wall, and stop. The glossary will translate the file-format words when a listing gets obscure. The blog and the directory cover the hosted tools you can keep using for everything that does not need to stay on the disk.
