Start with your workload, check what your computer can support, then compare offers. You may already have enough hardware to get started.
Practical local AI guides
- Try your first local AI chat on a supported computer you already own.
- Test your local AI workload before you buy and record what needs improving.
- Understand VRAM vs unified memory and build a realistic memory budget.
- Check a used GPU before buying, then test it safely after delivery.
1. What do you want to do?
- Chat, write or code for yourself: Try your chosen model on your current computer first. Note what you want to improve: answer quality, response speed or how much material it can read at once.
- Work with long documents or coding agents: Test the context length you actually need. Context is the material the model handles in a request; longer context can require more working memory.
- Run several requests at once: Check performance under that load. A model that fits for one conversation may need more memory for parallel requests. How concurrency affects memory
- Generate images or video, or fine-tune models: Start with your exact application’s hardware requirements. A language-model memory estimate alone isn’t enough to choose a machine for these jobs.
No model picked yet? Try a small local model before buying. Use a few tasks you care about to decide what “good enough” means.
2. Set a whole-system budget
Include the card or computer, any needed power supply and cooling, delivery and taxes. For an upgrade, check your case clearance, available slots and power connectors before comparing prices.
Then choose the closest starting point. A category may have no qualifying offers right now; it is worth waiting for suitable hardware rather than choosing an unrelated listing.
I have a compatible desktop and want to compare GPU upgrades
Start with the 24GB GPU offers. This is a comparison category, not a minimum requirement for local AI. Check whether your chosen model and settings fit with working memory to spare before spending more.
Watch for: used-card condition, power requirements and software compatibility. Cards with the same memory capacity can behave very differently.
Look for the 24GB GPU category when browsing the market board.
My workload needs more GPU memory
Compare the 32GB+ GPU offers. Use this route when measurements or reliable tests of your intended setup show that a smaller card is too limiting. More memory can make room for a larger workload; it doesn’t by itself promise faster answers.
Watch for: complete system cost and evidence from the same model, quantization, context and software you plan to use.
Look for the 32GB+ GPU category when browsing the market board.
I want a complete machine with a large memory pool
Compare high-memory systems. Check how much memory the GPU can actually use, which software supports the machine, and performance on your workload.
Watch for: memory labels. On Apple silicon, CPU and GPU share unified memory. System memory also serves the operating system and other apps, so an advertised unified-memory capacity isn’t a like-for-like dedicated VRAM figure. Apple’s explanation
Look for the high-memory systems category when browsing the market board.
3. Check fit before you buy
A model’s download size is only part of the memory budget. Allow for its working memory, context cache and other running software. Quantization stores model data at lower precision to reduce memory needs, with trade-offs that depend on the method, model and task. Check the exact version you intend to use. Quantization explained · Context-cache memory
Before choosing an offer:
- Confirm your application supports the exact GPU, operating system and driver combination. Ollama’s hardware list is one example; check your own app too.
- For a used card, ask for evidence it works under load, confirm the exact model and read the seller’s return terms.
- Check that card manufacturer’s dimensions, cooling and power requirements. NVIDIA’s RTX 3090 specifications illustrate the checks; partner cards can differ.
- Compare the delivered total and listing condition. Recheck price and availability with the seller.
Ready to compare? Browse the market board, then verify the exact model, condition and delivered total with the seller. Keep your workload and total budget beside you.
For local AI questions, corrections or a build you want to share, visit the HighVRAM community page for posting tips and copyable templates.