Can I Run AI Locally on My PC, GPU, or Mac?

Can I Run AI (canirunai.app) is a free website that estimates which open-weight LLMs your computer can run locally. Pick or detect your GPU, Mac or CPU and choose a quantization and context length: every model in the catalog then gets a fit verdict and a memory estimate, and every model that can run also gets an expected speed range and, where available, ready-to-copy run commands. It works in the browser with no account or install, and it never downloads or runs a model itself.

You can run AI locally when the model weights, context KV cache, and runtime overhead fit within your practical GPU VRAM, Mac unified memory, or a workable combination of accelerator and system memory. Use the checker above to compare those requirements with your hardware.

That depends on your usable VRAM or unified memory, system RAM, memory bandwidth, selected quantization, and context length. Confirm your hardware, then filter the results by chat, coding, reasoning, vision, fit, provider, speed, memory, or popularity.

There is no single VRAM requirement for every local LLM. Parameter count, architecture, quantization, context length, KV cache, and runtime overhead all matter. The LLM VRAM calculator estimates these pieces together and shows whether the result is comfortable, tight, offloaded, or too large.

Many laptops can run smaller or more heavily quantized local AI models, but the useful choice depends on GPU VRAM or shared memory, system RAM, and bandwidth. Select the closest laptop GPU or enter custom values to get a more relevant estimate.

Apple silicon Macs can use unified memory for local LLM inference. Select your Mac chip and memory size so the checker can account for practical available memory, bandwidth, model weights, context, and runtime overhead.

They describe how model weights are stored. Lower-bit quantization usually reduces memory use and can improve speed with some quality tradeoff. BF16 preserves original precision, while FP8 and NVFP4 require supported NVIDIA hardware and a matching model artifact.

They are planning estimates, not measurements. Memory is the sum of the model weights (parameters × the bits of the chosen quantization), the KV cache, sized from the model’s layers and attention heads for the selected context, and runtime overhead, and it is compared with the memory your hardware can actually use rather than the advertised total. Speed is a tokens-per-second range derived from memory bandwidth, because every generated token reads the active weights and the KV cache. Drivers, runtime versions, cooling, power limits and background apps can push real results outside that range, so test a tight fit before relying on it.

No. Tokens per second are estimated from public hardware and model specifications. Drivers, runtime versions, cooling, context length, background load, and the exact model build can change real performance.

The catalog lists 55 open-weight text-generation models from 12 publishers. Among them are Llama, Qwen, Gemma, DeepSeek R1 distills, Mistral, Phi and GPT-OSS, and each model lists its parameter count, native context window, license and source link. There is no fixed schedule: each refresh adds new models and re-checks existing entries against the publishers’ official model cards. The version in use (2026.10.19-1, October 19, 2026) is shown below the model list, and a share link keeps the catalog it was created with.

Automatic hardware detection runs in your browser, and raw GPU adapter details are not uploaded for detection. If the detected profile looks wrong, you can select a preset or enter custom memory and bandwidth values instead.