The beginner recommendation
Start with software before hardware. Install Ollama or LM Studio, load a small model, and test a real task. That one test will tell you more than a spec sheet.
Use this first model path:
ollama run llama3.2:3b
If that is slow, step down to a 1B class model. If it is smooth, try a 4B or 7B class model. Do not start with a large model because it sounds more capable.
Setup tiers by use case
| Setup tier | Planning range | Good for | Main limit |
|---|---|---|---|
| Current PC | $0 | Learning, tiny models, basic tests | May be slow or memory-limited |
| Basic upgrade | $300 to $800 | More RAM, SSD space, used office desktop path | Weak GPU or old CPU may remain |
| Practical beginner desktop | $900 to $1,800 | Small to medium local models, Ollama, LM Studio | Not ideal for large models |
| Strong hobbyist setup | $1,800 to $3,500 | 7B to 14B class models, coding, Open WebUI | Still limited by VRAM |
| Workstation | $3,500+ | larger models, heavier documents, small-team serving | Expensive and maintenance-heavy |
These ranges assume U.S. retail or used-market planning. They are not guarantees, and they do not include every monitor, backup drive, license, or optional cloud API cost.
What hardware matters most
RAM
RAM decides whether the machine has enough general memory for the operating system, apps, browser, local AI tool, and model workload. For local AI, 16 GB is a starting point, 32 GB is better for everyday use, and 64 GB gives room for heavier experiments.
If you are buying a Windows desktop, choose upgradeable RAM when possible. If you are buying a laptop or Mac, assume memory may be fixed after purchase.
VRAM or unified memory
VRAM is the memory on a dedicated GPU. More VRAM lets more model work stay on the GPU, which usually improves speed. A 4 GB GPU can help small models, 8 GB is more useful, 12 GB is a better budget target, and 16 GB to 24 GB gives more serious room.
Apple Silicon Macs use unified memory instead of separate system RAM and VRAM. That can work well for local AI, but the memory is shared with the whole system. For local AI on a Mac, 32 GB unified memory is a more practical target than 8 GB or 16 GB if you are buying specifically for this use.
GPU acceleration
NVIDIA GPUs remain a common path for local AI on Windows and Linux because many tools and libraries support CUDA. Ollama also documents support paths for NVIDIA, AMD Radeon, Apple Metal, and experimental Vulkan, with platform caveats. Check current support before buying around an older or unusual GPU.
Do not buy a GPU by name alone. VRAM, driver support, power limits, cooling, and desktop versus laptop versions matter.
SSD storage
Model files are large. A beginner can fill a small SSD by downloading several model families. Use at least a 1 TB SSD if local AI will be more than a weekend test. A 2 TB SSD is useful if you keep models, documents, embeddings, code projects, and Docker data.
CPU and cooling
CPU matters more when you run CPU-only inference or when the system coordinates tools, documents, and Docker. Cooling matters because local models can keep a laptop or desktop under sustained load. A thin laptop may technically run a model but throttle quickly.
What you can run by goal
For simple chat and summaries, a 16 GB to 32 GB machine can be enough if you choose small models. For coding help, use a coding model that fits your hardware instead of only chasing larger general models. For document workflows, memory and context length matter more, so 32 GB to 64 GB is more comfortable.
For Open WebUI with one user, a modest desktop is fine after Ollama works. For small-team Open WebUI, choose a stable host with backups, enough RAM, and a clear admin owner. For multiple users or heavy documents, do not treat a personal laptop as a server.
Windows, Mac, or Linux
Windows is the easiest mainstream path for many beginners because Ollama, LM Studio, Docker Desktop, and common GPU drivers are available. It is also the most common path for gaming desktops with NVIDIA GPUs.
macOS is strong on Apple Silicon because unified memory can be efficient and the app experience is simple. The downside is that memory and storage are often expensive and fixed after purchase.
Linux is flexible for server-style setups, NVIDIA workstations, and remote hosts. It also expects more comfort with drivers, permissions, services, and command-line troubleshooting.
Before you buy hardware
Ask these questions first:
- Have you run one small local model on your current machine?
- Is your bottleneck RAM, VRAM, disk space, CPU speed, or the tool setup?
- Do you need local AI for privacy or offline use, or only curiosity?
- Is your real task chat, coding, summaries, documents, automation, or team access?
- Would a cloud API be cheaper for occasional high-quality answers?
- Can you upgrade the current PC instead of replacing it?
If you cannot answer those yet, buy nothing. Run a small model first.
Common mistakes
The biggest mistake is buying for vague "AI" branding instead of a concrete local model goal. The second is buying too little memory. The third is buying a laptop that looks powerful but has limited VRAM, soldered RAM, weak cooling, or a tiny SSD.
Do not assume local AI is automatically cheaper than a subscription. A $2,000 PC is hard to justify if your real usage is a few prompts per week. Local hardware makes more sense when privacy, offline use, repeated experiments, or control matter.
Bottom line
For beginners, the best AI PC setup is a practical test machine, not a trophy build. Start on your current computer, move toward 32 GB RAM and a 1 TB SSD if local AI becomes useful, and buy GPU power only after you know the model size and workflow you need.