Can my PC run AI?

Every model

The memory each model needs at 4-bit compression, the usual download, with an 8,192-token conversation (about 6,000 words). A model runs fast when all of it fits in the graphics card’s memory, with about 1.2 GB to spare for the driver and its working space.

Up to 6 GB: any recent PC

6 to 14 GB: a graphics card with 8 to 16 GB

14 to 30 GB: a graphics card with 16 to 32 GB

30 to 100 GB: two cards, a Ryzen AI Max, or a card and plenty of system memory

Over 100 GB: a workstation or a server

What each part of a PC does for AI

Best GPU for local AI compares every graphics card’s tokens per second, with a page for each card listing the models it runs. For image and video models — Stable Diffusion, FLUX and Wan — see Can my PC run Stable Diffusion?

• Graphics card. The fastest place to run a model: its memory is read five to thirty times faster than system memory. What matters most is how much memory it has, because the model has to fit, then that memory’s speed. Two cards, even two different ones, pool their memory to hold a bigger model; they do not make one that already fits any faster.
• Graphics built into the processor. A laptop’s Radeon 780M or Arc 140V uses system memory, so it writes at about the speed that memory allows, but reads a prompt several times faster than the processor alone.
• Processors with fast shared memory. The Ryzen AI Max+ 395, Apple’s Max and Ultra chips and NVIDIA’s DGX Spark give their graphics up to 128 GB (512 GB on a Mac Studio) at 250 to 800 GB/s: slower than a big graphics card, but able to hold models no gaming card can.
• System memory. How much there is decides which models fit at all without a graphics card, or beside one; its speed and channels decide how fast the part held there runs. Two channels are twice as fast as one, and a server’s twelve, six times.
• Processor. It matters little for writing speed, which memory sets, except a small one (a mini PC’s N100, a Raspberry Pi) that cannot keep its memory busy. Without a graphics card it sets how fast a prompt is read.
• Drive. It holds the model file and loads it: a 40 GB model loads in about 8 seconds from a PCIe 4.0 SSD and about 4 minutes from a hard drive. A model too big for memory can be read from the drive as it runs, which is slow even from the fastest drives; only very large mixture-of-experts models, which use a small part of themselves for each token, stay near a token a second.
• Neural processors (NPUs). The NPUs in Copilot+ laptops run small models in some apps, such as Microsoft’s Foundry Local, AMD’s Lemonade and Intel’s OpenVINO, but they share the laptop’s memory, so they write no faster than its built-in graphics. The most popular apps run models on the graphics or the processor instead.

How it is worked out

Writing an answer is limited by memory rather than arithmetic: for every word-piece (token) a model writes, the computer reads every weight that token uses once. So the speed is how fast the memory holding the model can be read, divided by how much is read for each token: the model’s weights at the chosen compression, plus the conversation so far. How much of its memory speed each graphics card reaches is measured, not assumed: it comes from the card’s own result in llama.cpp’s benchmark, which LM Studio, Ollama and Jan are built on. A model that does not fit in the graphics card’s memory spills into system memory, five to fifteen times slower, and one that does not fit in memory at all can be read from the drive as it runs, slower again; the estimate counts each part at its own speed, and a processor with few threads cannot use all of fast memory. Graphics built into a processor use system memory, at its speed. Reading a prompt is limited by arithmetic, so it follows the measured prompt speed.

• Graphics cards’ speeds: llama.cpp’s community scoreboards on GitHub for NVIDIA (CUDA), Vulkan (AMD and Intel cards, and graphics built into processors), AMD (ROCm) and Apple Silicon: Llama 2 7B at 4 bits on each, writing and reading a prompt. Parts without a result there are estimated from their nearest relative.
• Checks against larger and newer models: Qwen3.5 27B on an RTX 3090 and gpt-oss 20B, Qwen3 14B and 32B on an RTX PRO 4500 (the same scoreboards), Meta’s own Muse Glimmer 30B figure on an RTX 5090, a published log of 51 models on a Ryzen AI Max+ 395, processors from an Intel N100 mini PC to EPYC servers running DeepSeek R1, and DeepSeek R1 read partly from an NVMe drive. The estimates fall within about a fifth of those measurements, which is why each is shown as a range.
• The models: each one’s own Hugging Face page and configuration, linked from its page here.

Running a model

LM Studio, Ollama and Jan are free and run every model here: search for the model’s name and choose a file of the size shown for your PC. Q4_K_M, about 4.8 bits a weight, is the usual choice: close to the original, a third of its size. A longer conversation needs more memory, so if a model only just fits, a shorter one, or the next size down, keeps it on the graphics card. Laptop graphics run slower than the desktop cards of the same name, by how much depends on the laptop.

Guides: Ollama vs LM Studio, which GGUF file to download, Ollama using the CPU instead of the GPU, LM Studio not using the GPU and shared GPU memory. And what tokens per second look like.

Models were read from their makers’ pages on 8 November 2026. “Open” here means the weights can be downloaded and run; licences differ, and each model’s page names its own. Bottleneck is not connected with any of these models or their makers.