CanIRun.ai: Is Your PC or Mac Ready for Local AI?

CanIRun.ai: A Solution to Evaluate Your Machine's AI Compatibility

CanIRun.ai is an online tool that allows users to determine whether their computer, whether a PC or a Mac, is capable of running artificial intelligence models locally. By analyzing your machine's hardware configuration, it indicates which AI models, from a selection of open-weights LLMs such as Llama, Qwen, Gemma, and Mistral, can be executed with a certain level of comfort.

On machines equipped with Apple Silicon, the tool treats unified memory as a single VRAM, but only utilizes about 75% for the AI model, which can limit performance, especially on devices like the Mac M1 16GB.

The entire process takes place in the browser. CanIRun.ai identifies the GPU/CPU components and briefly assesses their performance to determine compatibility, without ever sending your data to an external server.

A Tool to Simplify AI Model Selection

For many, knowing whether their computer can run AI locally is a real headache. Fortunately, a new free tool is here to simplify this task. CanIRun.ai scans your hardware and translates technical jargon into a simple and understandable verdict.

In recent months, tools for running AI models locally have proliferated, but one question remains: what can your machine actually support? The CanIRun.ai site detects your hardware configuration and tells you which AI models can be run locally, as well as the expected performance metrics, such as VRAM usage and generation speed.

In practical terms, the site presents itself as a catalog of models (Llama, Qwen, Gemma, Mistral, etc.), each accompanied by a clear indication of its compatibility with your machine, ranging from very comfortable to completely out of reach. The goal is to make intimidating technical concepts (VRAM, quantization, billions of parameters) accessible to a non-specialist audience.

Practical Use of CanIRun.ai

The CanIRun.ai interface automatically displays a summary of your hardware configuration, followed by a list of models annotated with a sort of report card:

• "Comfortable" models
• Models that work but are "limited"
• "Too heavy" models to avoid

To use this tool, simply:

• Open CanIRun.ai in a recent browser.
• Allow the site to detect your configuration and quickly verify that the displayed information (chip, memory, GPU) is correct.
• Browse the list of models and identify those classified in the most comfortable categories, based on your usage (general chat, coding, etc.).
• Install one of these models via a tool like Jan, Ollama, or LM Studio to get started with local AI without turning your machine into a radiator.

CanIRun.ai focuses solely on "self-hostable" models, meaning those whose weights can be downloaded for local use on your GPU or Apple Silicon chip. The tool relies on a catalog of popular open-source or open-weights models, ranging from small models of 1 to 2 billion parameters to very large models of over 100 billion parameters. It does not list large purely proprietary cloud models like ChatGPT, Claude, or Gemini in their online versions, as they cannot be run locally on a standard PC.

How CanIRun.ai Works on Mac

With a 2020 MacBook Air (with M1 chip and 16GB), CanIRun.ai operates consistently, although it sometimes detects an M1 Pro chip instead of M1. When it comes to an Apple Silicon machine with unified memory, the tool displays the total amount of available memory as if the machine had 16GB of VRAM.

The RAM box is then grayed out with a note about unified memory, indicating that it is not two separate blocks, but a single pool of 16GB shared between the CPU, the GPU, and the rest of the system.

Given that macOS and other applications already consume several gigabytes, the available margin for a local model is reduced. These figures should therefore be considered as rough estimates rather than guarantees of performance. For a Mac M1 16GB, small and medium models can be run, but larger models may struggle. However, the creators of the tool highlight a specific advantage of Apple Silicon: a model can use up to about 75% of the total memory, unlike a dedicated graphics card that is strictly limited to its VRAM.

Technical Aspects of CanIRun.ai

Technically, as soon as you open the site, the page thoroughly queries what the browser can reveal about your hardware. CanIRun.ai creates a small graphical context in the background to retrieve the exact name of your GPU or Apple Silicon chip, then cross-references this information with an internal database of graphics cards and processors to find their actual specifications (amount of memory, bandwidth, architecture, etc.).

At the same time, it uses the system information exposed by the browser to estimate RAM and the number of processor cores, and performs a brief performance test to assess single-core power. All this data serves to reconstruct a hardware profile sufficiently precise to determine which models can be loaded into memory and at what speed they can operate, all directly in the browser and without sending these measurements to a server.

To refine its verdict, the tool also takes into account how the AI is compressed (for example, in 4, 8, or 16 bits, known as quantization).

However, this tool is not infallible. Detection depends on what the browser allows to be exposed, and some environments may mask the graphics card name, leading to less accurate estimates. This is why the authors offer a mode where users can manually enter their configuration. Similarly, compatibility scores remain estimates: they provide a reliable order of magnitude for choosing a model or considering a future GPU purchase, but do not replace a complete benchmark for specific use.