LLM hardware requirements: what you actually need
The most common question in local AI is some version of "can my machine run this?" The answer comes down to one number: how much memory the model can live in. This page is the sizing chart we hand every new member, with the reasoning behind it.
The one rule that explains everything
A quantized model (q4, the standard sweet spot of quality versus size) needs roughly 0.6 GB of memory per billion parameters, plus 2 to 4 GB of headroom for context. An 8B model wants about 5 GB. A 14B model wants about 9 GB. A 32B model wants about 20 GB. Where that memory lives determines your speed: GPU VRAM is fast, Apple unified memory is nearly as good, and CPU RAM works but is several times slower.
The sizing chart
| Your memory | Model class | Feels like | Good for |
|---|---|---|---|
| 8 GB RAM, no GPU | 3B | Quick, modest depth | Autocomplete, tagging, simple Q&A, learning the tooling |
| 16 GB RAM, no GPU | 8B | Capable, slow on CPU | Summaries, drafts, light document search |
| 12 GB VRAM or 32 GB RAM | 14B | Fast and truly useful | Coding help, document work, daily agent workloads |
| 16-24 GB VRAM or 48 GB unified | 32B | Where local gets seriously good | Most dev workflows, strong RAG and agents |
| 48+ GB VRAM or 64+ unified | 70B | Near frontier, always on | Hard reasoning, long context, serving a household |
What to buy at each budget
Under $400: a used mini-PC. A renewed Ryzen box with 32 GB of RAM runs 8B models and teaches you the whole stack before you spend real money. This is the right first purchase for most people, and it is the one we most often talk members into instead of a GPU they do not need yet.
$900 to $1,600: a 16 GB GPU workstation. The workhorse tier. Runs 14B models fast and 32B models with careful quantization. If you do real work with local AI, this is where the price-to-capability curve is kindest.
$1,000 to $2,400: Apple silicon. A Mac mini or Studio with 32 to 64 GB of unified memory runs large models silently and sips power. The GPU can address roughly two thirds of unified memory for weights, so a 48 GB Mac is effectively a 32B-class machine. The best choice for an always-on agent box on a desk.
The buying rules
- Buy VRAM first, RAM second, CPU last. A 12 GB GPU outruns 64 GB of CPU-only RAM. CPU speed barely matters for inference.
- Capacity beats speed. A slower card with 16 GB is more useful than a faster card with 8 GB, because it runs models the faster card simply cannot load.
- Already own a gaming PC? Start there. A 12 GB card from the last few generations is a legitimate 14B-class machine today, at a cost of zero dollars.
- Do the break-even math before buying anything. If you spend $150 a month on AI subscriptions, a $1,200 rig pays for itself in eight months. Our calculator does this with your numbers.
Memory prices are volatile in 2026 and specific model recommendations shift monthly as new releases land. We keep current picks and prices updated in the RunLLMsLocally Discord, and the free starter kit includes a script that detects your hardware and picks the right model automatically.