Which setup for which budget? 3 configs for a local LLM

Three ways to run a local LLM: an entry-level setup, the best value, and no limit. The model each one runs, PC or Mac, with marketplace links.

For a local LLM, a budget boils down to one thing: the memory available to the model (a graphics card's VRAM, or a Mac's unified memory). It sets the model size, so its quality; bandwidth then sets the speed. Here are three setups, from entry level to no limit, each as a PC and, where one exists, as a Mac.

This guide shows no prices: they change too fast. Each link opens a search on your country's store, or the product page with its dated price when an offer is available. These are affiliate links: they change neither our recommendations nor the calculations.

What each setup runs

Setup 1: Budget, a good LLM without overspending

An RTX 3060 12 GB is enough for 8 to 14 billion parameter models, which handle everyday uses well (writing, summaries, simple code). If you can step up, the RTX 5060 Ti 16 GB adds context headroom. On the Mac side, a Mac mini M4 with 16 GB is the simplest and quietest entry point, but its memory is shared with the system, which quickly limits the model size. Avoid 8 GB cards (see 12, 16 or 24 GB of VRAM).

Setup 2: the best value

The 24 GB tier is where you move to 27 to 32 billion parameter models, a clear jump in quality. The RTX 3090 offers 24 GB of VRAM; it is a previous-generation card, often found second-hand (check the condition and the seller's warranty). On the Mac side, a Mac mini M4 Pro with 48 GB holds bigger models, but generates slower than an NVIDIA card of the same capacity.

Setup 3: no budget limit

The RTX 5090 (32 GB) is the consumer card in our catalogue with the most memory: it runs 32 billion parameter models with plenty of context headroom. But it does not hold 70 billion parameter models. For those, the unified memory of a Mac Studio M3 Ultra with 256 GB changes the game: it loads models that no single card holds, at the cost of slower generation than an NVIDIA card on models that fit in VRAM.

Speed, setup by setup

Qwen3 32B fits a 24 GB card; Llama 3.3 70B exceeds the VRAM of every consumer card. These speeds are bandwidth estimates (low to medium confidence): when the model does not fit, part of it runs from system RAM and speed drops.

The rest of the PC

For a PC build, the graphics card does all the work: the processor matters little as long as the model fits in VRAM (see CPU vs GPU). Plan for RAM at least equal to the VRAM, to load the model and offload layers if needed (32 GB for the Budget setup, 64 GB beyond; check the memory type your motherboard accepts). Model files are large: a 70 billion parameter model needs about 44.5 GB of memory (estimate), hence a 2 TB NVMe SSD. For the power supply, start from the wattage recommended by the card's maker and keep some headroom: an RTX 3090 or 5090 draws several hundred watts (see their pages).

Next steps

For a specific card, see which GPU to buy and Mac or PC for a local LLM. For your exact case, enter your hardware in the configurator: it computes the memory needed, the estimated speed and the compatible models.

EstimateData updated 39 days agoSource: Run LLM Locally engine