Can it run local AI?

Hardware check only — this page runs no AI and no model. It estimates whether your machine could run one, and how fast, from measured specs. Nothing is downloaded, generated, or sent anywhere.

Which open models fit on this PC and how fast they'll answer — worked out from what the benchmark actually measured (VRAM, memory bandwidth, compute), not from a spec sheet. Running a language model is a memory problem first: the weights must fit, and every generated token re-reads all of them, so bandwidth sets the speed. Edit any number to explore a different card or a RAM upgrade.

Your rig

Green = measured by this site on this machine · blue = from the part's spec sheet · amber = estimated. Anything can be overridden.

No GPU detected yet — run the benchmark once, or type your card's numbers below.

VRAM (GB)

GPU memory bandwidth (GB/s)

GPU compute (FP32 TFLOPS)

System RAM (GB)

CPU memory bandwidth (GB/s)

Run the benchmark to fill these in from measurements →

Language models

Tokens per second while generating (a comfortable reading pace is ~10; 30+ feels instant). "Fits in VRAM" runs entirely on the GPU; "split" spills into system RAM and slows to the RAM's pace for the spilled part; "too big" won't load at this quantisation.

Model	Size	Fit	Est. speed	How it feels

Image generation

Diffusion models are compute-bound rather than bandwidth-bound: the VRAM floor decides whether they load, the TFLOPS decide how long one image takes.

Model	VRAM needed	Fit	Est. per image

How the estimate works. Model size = parameters × bytes per parameter for the chosen quantisation (4-bit ≈ 0.57 B, 8-bit ≈ 1.06 B, FP16 = 2 B) plus a working-memory allowance for the context window. Generation speed ≈ memory bandwidth ÷ bytes read per token × an efficiency factor (~55 % on a GPU, ~70 % of effective multi-channel bandwidth on a CPU); for mixture-of-experts models only the active parameters are read per token. When a model spills into system RAM, the spilled fraction runs at the RAM's bandwidth and dominates. Image times = model cost ÷ FP32 TFLOPS. These are order-of-magnitude estimates with a ±25 % band — real numbers depend on the runtime (llama.cpp, Ollama, LM Studio, vLLM), context length, drivers and thermals. Nothing here is uploaded; the page reads the last benchmark stored in this browser.