About CanRun

CanRun answers one question: can this PC or Mac run this local LLM, and how well? Pick a graphics card, an Apple chip or a CPU-only setup, and CanRun shows which open models fit, which quant to download, how much memory each one needs and roughly how fast it will generate text.

What you get

• A verdict for every model: Runs great, Runs well, Runs slowly or Won't run, based on whether the model fits in memory and how fast it is likely to run.
• The memory it needs, split into the model weights, the KV cache for your context length and a small working buffer.
• A speed range, never a single number. Every speed comes with a confidence label that says how sure we are.
• Ready-to-run commands for llama.cpp and Ollama, matched to the context length you pick.

The calculator covers any combination, and every GPU, model and model-on-GPU pair also has its own page.

Why CanRun exists

Plenty of calculators tell you how much memory a model needs, but few say how fast it will actually run on your hardware, or how far that number can be trusted. CanRun was built to answer both: every speed estimate is checked against real benchmarks, and how far it has been off is published rather than hidden. It is written in English and Korean, because there was hardly any tool like it for Korean readers.

Where the numbers come from

Model sizes come from the actual GGUF files on Hugging Face, and memory figures from each model's published configuration. GPU and chip specs come from manufacturer pages. Speed estimates start from memory bandwidth and are calibrated against public benchmarks. The full method, how accurate it has been and where it falls short are on the How we calculate page.

Since October 2026 we also run models ourselves. The first two were Llama 3.1 8B and Gemma 3 12B on a GeForce RTX 5060 Ti 16GB with Ollama, and the RTX 5060 Ti 16GB guide shows the PC, the script anyone can run to repeat the test and how the results compare with CanRun's estimates. Every measured speed on the site, ours or public, is labeled Measured and says where it came from.

How often it is updated

Model data is checked against Hugging Face every week, and new models are added once llama.cpp supports them. The date of the current data is at the bottom of every page. It is currently October 28, 2026. The Updates page lists what has changed on the site and the date each model was added.

How mistakes get fixed

When a number turns out to be wrong, we correct the data or the calculation and add a note to the Updates page. Changes from the weekly Hugging Face check go through automatic checks before they reach the site, and large or unusual changes are looked at by hand first.

If a number looks wrong or a model is missing, email [email protected] or open an issue in the CanRun repository on GitHub. GitHub issues are public, so use email for anything you would rather keep private. The contact page lists what helps us check a report.

Who runs CanRun

CanRun is an independent project run by one person, YOO, KING. It is not affiliated with any GPU maker or model publisher, and its verdicts are calculated from data, not from any commercial arrangement.