# Can I Run This LLM? Check What Local AI Models Your PC Can Run | Koso > Free local LLM compatibility checker. It detects your GPU, VRAM and RAM in the browser, measures every popular open model against your machine using real quantisation file sizes, and tells you which ones run, how fast, and the command to pull them. Get alerts when a new model that fits your PC ships. Free AI agent # Can I run this LLM? We read your graphics card straight from the browser, measure every popular open model against it using real download sizes, and tell you which ones run, how fast, and what to type to get them. Reading your hardware Want exact specs? Download the scanner Free · first run without sign-in · results in about 30 seconds ## How the answer is worked out Nothing here is a language model guessing at numbers. The fit, the context ceiling and the speed are arithmetic over real file sizes and real card specifications. The model only writes the explanation once the table is finished. ### Real sizes, not formulas A 7B at Q4\_K\_M is 4.68 GB, not the 3.5 GB that bits times parameters predicts. The sizes come from each model repository's actual file listing, refreshed daily, and on an 8 GB card that difference decides the answer. ### The KV cache is counted A model that fits on paper falls over at long context, because the attention cache grows with every token. That is why the answer is given per context length rather than as one number, and why the table has a max context column. ### Mixture-of-experts is handled A 30B-A3B needs the memory of a 30B and generates at roughly the speed of a 3B. For anyone with 16 to 24 GB that changes which model they should download, and no calculator keyed on parameter count alone can say it. ### Not fitting is not the same as not running Split across VRAM and system RAM a model runs at the harmonic mean of both bandwidths. Usually slow, often usable. Saying no there would be false, so the table says slow and shows you the number. ### The pipeline your browserWebGL + WebGPU ×2checked in7 days?truefalsedetect-agentcard → VRAM · bandwidthcatalog-agentlive models · real sizesfit-agentweights + KV + overheadadvisor-agentthe write-up · AIGPU tableVRAM · GB/sHugging Facerefreshed dailyprompt storeeditable, no deployyouranswercached 7 daysalert-agentweekly · only whatfits your machine The actual pipeline behind this page. On a phone, swipe the diagram sideways to follow it. ## Questions How does it know what graphics card I have? Your browser tells it. WebGL exposes the renderer string for the GPU it is drawing with, and WebGPU is asked twice, once preferring high performance and once preferring low power, because a laptop with both integrated and discrete graphics will often only admit to one of them at a time. Nothing is installed and nothing about your machine is sent anywhere except the specs needed to compute the answer. Why does it ask me to confirm my VRAM? Because no browser will tell a website how much video memory a card has. That is deliberate, since it would help fingerprint you. VRAM is matched from the card's model name instead, and several cards shipped in more than one memory size: the RTX 3060 came as 8 GB and 12 GB, the 4060 Ti as 8 GB and 16 GB. Guessing wrong changes the answer for every model in the 13B class, so it asks instead. What is the downloadable scanner? A small program that reads your real specs and sends them straight back to this page, so your numbers appear in the browser without copying anything. It exists because browsers will not report VRAM, cap reported RAM at 8 GB, and often hide a laptop's discrete GPU. It reads specs, makes one request to this site with the numbers it shows you, and exits; its full source is on the page next to the download. If Windows asks about an unrecognised file: More info, then Run anyway. Where does the list of models come from? The Hugging Face API, refreshed daily, filtered to models with real download numbers behind them and to publishers that are not bulk re-quantisation mirrors. Crucially the file sizes are the real ones read from each repository, not estimates from a formula. A 7B at Q4\_K\_M is 4.68 GB, not the 3.5 GB the arithmetic suggests, and on an 8 GB card that difference is the whole answer. How accurate are the tokens per second estimates? They are estimates, and the page says so. Local generation is limited by memory bandwidth for large models and by compute for small ones, so both limits are modelled and combined. Against published benchmarks the numbers land in the right band, but your actual speed depends on your quantisation, your context length, your inference engine and what else is using the GPU. Treat them as a band, not a benchmark. What is a mixture-of-experts model and why does it matter? A model like a 30B-A3B has 30 billion parameters but only activates about 3 billion for each token. It needs the memory of a 30B model and generates at roughly the speed of a 3B one. For anyone with 16 to 24 GB that is the single most useful fact on this page, because it means the largest model you can run is often far faster than you would expect. How is this different from Can You Run It (CYRI)? CYRI answers whether a game will run; this answers which local AI models will run, which is a memory question rather than a frame rate one. The mechanics are deliberately similar: browser detection first, and when that is not enough, a downloadable scanner that reads the machine directly and reports back to the page on its own, so the result appears without copying anything. The difference is transparency: our scanner's full source is printed on the page, it sends exactly one request with the numbers it shows you, and the fit itself is computed from real model file sizes rather than a requirements database. What are the new model alerts? Once a week, if a newly released model fits your machine, you get one email naming it, the speed you should expect, and the command to pull it. If nothing new fits that week, you get nothing. Your hardware profile is stored with your address so the filtering actually happens: someone on an 8 GB card is never told about a 70B. One click unsubscribes, no login needed. Is it really free? Yes. The first run needs no sign-in. After that it asks for a name and an email, no password and no credit card, which is also how you get your run history. Shared links and cached results are always free and never ask for anything. ### Built by Koso We build multi-agent systems for companies. This one is free and public because showing the work is a better pitch than describing it. [Book a call ](/book)