What AI can my computer run?

Because without a graphics card the speed is set by memory bandwidth, not by the processor. Generating a word requires reading the model's active weights out of memory, so the ceiling is bandwidth divided by the size of those weights, and no processor makes that faster. Commodity DDR3 moves about 21 GB/s, DDR4 about 50 and DDR5 about 90. The same model on a 2012 machine and a 2024 one differs by roughly four times, which is the difference between an answer in seconds and an answer you wait out. Age also stands in for the instruction set: Ollama's official builds expect AVX, and machines from before about 2013, including AMD's FX series, may not start them at all. To check, open Task Manager on Windows and look at Performance, then Memory, which names the generation. On Linux use lscpu and dmidecode. Apple Silicon is not asked, because it is all recent and its memory is fast.

Not reliably. Browsers deliberately hide much of that information for privacy, and they do not expose dependable VRAM figures. This release asks for the few values that change the result rather than pretending it detected facts the browser cannot see. A separate, consent-based local detector can be added later.

Probably, but not on the path most people assume. Ollama's supported list, published at docs.ollama.com/gpu, covers the RX 9000 and 7000 series and, from RDNA 2, only the RX 6950, 6900 and 6800. An RX 6700 is not on it, and neither is an RX 5000, a Vega or the very common RX 580. Those cards are not excluded, because Ollama also provides support through Vulkan, and on Windows the AMD driver already ships Vulkan so there is nothing extra to install. Linux may need Vulkan drivers from Mesa or AMD. The practical difference is that ROCm is the supported acceleration path and Vulkan is the fallback, so the tool now asks which Radeon you have and names the one you will actually be using rather than offering both and leaving you to guess. If you are not sure, it assumes Vulkan, because that is the cautious direction.

Yes. The calculator includes Linux with NVIDIA, Linux with AMD, Linux CPU-only and Windows with AMD. Memory fit can be calculated from VRAM and RAM, but actual AMD acceleration still depends on whether the specific card and installed driver are supported by ROCm or Vulkan. The result keeps that driver check visible.

No. The model catalogue arrives with the page and every calculation happens locally in your browser. Your selections are not submitted, saved or placed in a URL. The page also carries a Content Security Policy that blocks the browser connection methods normally used to send them.

Four-bit quantisation substantially reduces the memory required to run a model locally while usually preserving enough quality for everyday work. The calculator names the exact Q4_K_M variant so its memory estimate refers to a real runnable configuration rather than a model family in the abstract.

No. Fast, comfortable and slower are calculated fit classes based on where the model can be held: fast GPU or unified memory, partly offloaded system memory, or CPU-only memory. The tool does not invent a tokens-per-second figure for hardware it has not measured.

Your operating system, browser and AI runtime need memory too. Treating every advertised gigabyte as available to model weights produces recommendations that load and then fail during a long conversation, so the calculator keeps a conservative working reserve.

Both run models on your own machine, and each result names the same model in both. Ollama is a command-line runner for macOS, Windows and Linux: the command shown downloads the named model from Ollama's catalogue and starts it. LM Studio is a desktop application with a search box and no command line, which suits most people better if a terminal is unfamiliar. Installation is your choice either way: this page copies a command or a model name and never executes anything on your computer.

On a 16GB Apple Silicon Mac, a 4-bit model around 8B parameters is usually the comfortable tier, while some 12B and 14B models may fit with less headroom. A Windows or Linux PC also depends on its GPU VRAM, so the calculator separates graphics memory from ordinary RAM before making a recommendation.

Small 3B and 4B models can fit inside roughly 4GB to 6GB, capable 7B and 8B models are more comfortable around 8GB, and 14B models generally want about 12GB or more for a fully accelerated 4-bit setup. Context length, runtime overhead and system use change the real requirement, which is why the calculator includes a safety reserve.