What you run local AI on. For most models the single most important number is memory: a model has to fit before anything else matters. These pages focus on that, plus bandwidth and power, and stay deliberately vague on price, which dates fast.
Enter your GPU or memory and see which models fit, and at what quantisation.
Pick a model and see the machine to run it, at each quality level, with indicative prices.
Work out whether buying hardware actually costs less than paying for an API, assumptions included.
NVIDIA · Budget
A budget favourite for getting into local AI. The 12GB variant offers more memory than several pricier cards, which makes it a sensible, affordable entry point for running small and mid-sized models.
Intel · Budget
Intel's Alchemist flagship with 16GB of memory, a budget route to 16GB of VRAM for local inference outside the CUDA ecosystem.
NVIDIA · Budget
A budget-friendly way to get 16GB of VRAM. Its memory bandwidth is modest, so generation is slower than pricier cards, but the 16GB lets it run models that do not fit on cheaper 8GB and 12GB options.
NVIDIA · High-end
An Ada Lovelace consumer card that pairs 16GB of memory with a 256-bit bus, a popular value pick for 16GB of CUDA VRAM.
NVIDIA · High-end
A high-end Ada Lovelace consumer card with 16GB of memory, a strong all-rounder for local models that fit within 16GB.
NVIDIA · High-end
A current-generation card with 16GB of GDDR7 and strong bandwidth. Fast and capable for models up to the mid teens of billions of parameters, though its 16GB limits it against the 24GB and 32GB cards for larger models.
AMD · High-end
An RDNA 3 consumer card with 20GB of memory on a 320-bit bus, offering more VRAM and bandwidth than similarly-priced 16GB NVIDIA cards.
AMD · High-end
AMD's high-end consumer card with 24GB of VRAM, and the main alternative to NVIDIA for local AI on a budget for memory. Support through ROCm has improved, though NVIDIA's CUDA ecosystem remains smoother for most tools.
NVIDIA · Mid-range
A previous-generation card that remains one of the best value choices for local AI, thanks to 24GB of VRAM at used-market prices. Slower than current cards, but the 24GB is what matters most for fitting models.
NVIDIA · Enthusiast
The previous-generation consumer flagship, now succeeded by the RTX 5090 but still an excellent local AI card, with 24GB of fast GDDR6X memory and strong compute. Enough to run capable models at good quantisations, though 24GB sets a real ceiling on model size.
NVIDIA · Enthusiast
The current consumer flagship, and a major step up for local AI thanks to 32GB of fast GDDR7 memory and enormous bandwidth. The extra 8GB over the 4090 lifts the ceiling on model size, and the bandwidth speeds up generation noticeably.
NVIDIA · Enthusiast
The Ada Lovelace successor to the A6000, with 48GB of ECC memory, faster tensor cores, and FP8 support, aimed at professional and prosumer local AI.
NVIDIA · Enthusiast
Ampere-generation professional workstation card with 48GB of ECC memory, a long-standing favourite for single-card local inference of large models.
NVIDIA · Server
A workstation card with 96GB of GDDR7, aimed at professionals and small teams who need to run large models on a single card. It offers the most memory available on a single non-data-centre GPU, at a professional price.
Apple · High-end
The M5 Max Mac Studio, announced August 2026, with up to 128GB of unified memory shared between CPU and GPU at a claimed 614GB/s. A capable, quiet desktop for running mid-to-large models, sitting below the M5 Ultra on both memory ceiling and bandwidth.
Apple · High-end
The laptop route into local AI. An M4 Max MacBook Pro pairs a capable GPU with a single pool of unified memory, so a large share of the machine's memory can hold a model. Configurable from 36GB up to 128GB, it runs mid-sized and, at the top configuration, large models on the move, though more slowly than a discrete GPU.
Apple · High-end
A previous-generation Apple Silicon desktop, now behind the M3 Ultra and M4 Max Studios, but still capable and good value used. Its unified memory is shared between CPU and GPU, so most of the 64GB is available to models, enough to run large models that would otherwise need multiple discrete GPUs, at low power.
Apple · Enthusiast
Apple's top desktop chip in the Mac Studio, with high-bandwidth unified memory shared between CPU and GPU. Its memory bandwidth is the highest of any Apple Silicon, which makes it one of the better Apple options for running larger models.
Apple · Enthusiast
Apple's top desktop chip in the Mac Studio, announced August 2026, with up to 512GB of unified memory at a claimed 1.2TB/s. That memory ceiling lets a single desktop hold very large open-weight models that would otherwise need a multi-GPU server, at a fraction of the power.
Apple · High-end
A Mac Studio built on the M4 Max, with up to 128GB of unified memory. It is a more affordable Apple route to running large models than the Ultra, with lower bandwidth but a very capable memory ceiling.
NVIDIA · High-end
A small desktop AI computer built on NVIDIA's GB10 Grace Blackwell chip, with 128GB of unified memory. It brings NVIDIA's software stack to a compact machine aimed at running and prototyping large models locally.
Apple · Mid-range
The refreshed Mac mini, announced August 2026, in its local-AI-relevant M5 Pro configuration with up to 64GB of unified memory. A small, low-power, affordable route into Apple Silicon for running mid-sized models. A base M6 chip, which Apple describes as its first 2-nanometre part, is also offered lower in the range.
Apple · Mid-range
A small, affordable desktop with up to 64GB of unified memory, and one of the cheapest ways into Apple Silicon for local AI. Its bandwidth is modest, but the unified memory lets it run models that need more than a mid-range card.
AMD · Mid-range
An APU with a large integrated GPU and up to 128GB of unified memory, sold in compact mini PCs. Up to 96GB of that memory can be assigned as VRAM, which lets it run models far larger than any single consumer discrete card, at low power.
NVIDIA · Budget
A 2016 Pascal data-centre accelerator with 24GB of memory, repurposed as a cheap high-VRAM card for homelab local inference.
NVIDIA · Server
A previous-generation data-centre accelerator with 80GB of HBM memory. Slower than the H100, but still highly capable and often more available or cheaper to rent, which keeps it common in professional and academic deployments.
NVIDIA · Server
A data-centre accelerator with 80GB of HBM memory and enormous bandwidth, built for training and high-throughput serving. It is the workhorse of professional AI infrastructure, and firmly organisational rather than home hardware.
NVIDIA · Server
A Hopper-architecture data-centre GPU, the first with 141GB of HBM3e at 4.8 TB/s, built for large-model training and high-throughput serving.
Not everyone who wants to run local AI on a Mac needs a Studio. A Mac mini with the higher memory option holds mid-sized models quietly and at low power, and costs far less than the large-model Apple Silicon machines.
For many people the best value in local AI is a used RTX 3090. Its 24GB runs models up to around 32B at a 4-bit quantisation, often for less than a new 16GB card, if you are comfortable buying second-hand.
A quiet, low-power route to running large models at home, built around a Mac Studio with plenty of unified memory. Slower generation than a discrete GPU, but able to hold models that a single consumer card cannot.
A first local AI machine built around a 16GB card, enough to run 7B to 14B models comfortably without spending flagship money. The sensible way to find out whether local AI suits you before committing more.
Image generation asks different things of a machine than a language model. A fast 16GB card handles SDXL and FLUX comfortably, and prioritises the bandwidth and compute that keep generation quick.
A compact, low-power route to running very large models, built around a Ryzen AI Max+ 395 mini PC with 128GB of unified memory. It holds models no single consumer GPU can, in exchange for slower generation.
A no-compromises single-card build around the RTX 5090. Its 32GB and high bandwidth run strong models up to around 32B quickly, with room for longer context than a 24GB card allows. Fast, capable, and expensive.
A single 48GB professional card runs 70B-class models on one GPU, with no multi-card complexity. It is expensive, but for a quiet, reliable workstation that holds large models, it is the cleanest option short of a server.