Can I run
CanIRun.ai helps you choose an open AI model for your own hardware. The hardware-matching site reads browser hints about GPU, RAM and CPU, matches them against a device database, and ranks models by estimated speed, memory headroom and quality. It covers chat, coding, image and video models.
Finding models that fit
You can visit the site in a desktop browser without an account or extension. If automatic detection misses your hardware, choose a device manually. The device list includes discrete GPUs, Apple chips, phones and Raspberry Pi hardware, allowing you to explore a machine before running a model on it.
The model list provides task, lab and license filters, plus sorting by factors such as parameter count, context length, speed and memory. Fit grades separate models that should run comfortably from tight fits or options too heavy for the selected machine. A model entry also lists available quantization formats, making it possible to compare smaller weights with higher-precision versions.
Understanding the estimates
The local AI documentation explains parameters, VRAM, quantization, context length and memory bandwidth. Quantization reduces the precision of model weights to save memory, with a quality tradeoff. The site commonly uses Q4_K_M for its approximate memory examples.
For Apple Silicon, unified memory replaces a separate GPU memory budget. Mixture-of-experts models still need space for the full model, even though only some experts are active for each token. Longer context also consumes more memory. These details help explain why an active-parameter count alone does not establish whether a model fits.
The rankings are estimates derived from hardware and model information. Use them to narrow a shortlist, then check the suggested format and memory requirements for the particular model and runtime you select.
Moving from a shortlist to local execution
A model page provides a runai command or guidance for using the same GGUF weights with Ollama or LM Studio. The publisher says runai chooses a quantization for the machine, downloads the weights, and starts local chat through llama.cpp.
CanIRun.ai therefore connects model discovery with the next setup step. Its documentation also explains GGUF, the weight-file format used by several local runtimes. License filters can help narrow the list for commercial or noncommercial work, but examine the selected model's license before deciding how to use its outputs.