CanIRun.ai

CanIRun.ai occupies a narrow but increasingly vital niche in the AI tool ecosystem: it tells you, before you hit download, whether your actual hardware can run a given local model. In an era where open-weight models like Llama 3, DeepSeek, and Qwen are multiplying rapidly, the gap between a model's listed requirements and your machine's real capacity is often wide and frustrating. This browser-based checker aims to close that gap using WebGPU APIs to inspect your GPU, CPU, and RAM on the fly, then cross-references those specs against a curated database of hundreds of models, complete with per-quantization VRAM estimates. The result is a compatibility grade from S to F, plus a breakdown of which quantized versions (Q2_K through F16) will fit and which will overflow your VRAM. It is not a benchmark tool, and it does not measure inference speed or throughput. It is a fit-checker, and for its intended audience, that is exactly what is needed.

The tool's standout strength is its instant, frictionless hardware detection. There is no software to install, no command-line incantations, and no manual spec entry. You land on the site, grant permission for browser hardware access, and within seconds the interface knows your GPU model, VRAM size, system RAM, and CPU architecture. For AI developers who cycle through multiple machines, or enthusiasts who want to quickly vet a new model before committing to a multi-gigabyte download, this immediacy is a genuine time-saver. The model database itself is thoughtfully organized. Beyond a simple search bar, you can filter by task (Chat, Code, Reasoning, Vision) and by provider (Meta, Alibaba, Mistral, Microsoft, etc.), which helps when you are looking for, say, a small code-specialized model that can run on a laptop GPU. Each model page shows not just a single VRAM number but a full table of quantization levels, each with its own memory footprint, so you can see exactly where the trade-off between quality and fit lands. For example, a user with an RTX 3060 12GB will see that Llama 3 8B at Q4_K_M fits comfortably, but Q8 is too large, and Q2_K might leave room for a larger context window.

The audience that benefits most is clear: local LLM enthusiasts who run models on personal hardware, AI developers evaluating deployment on edge devices or on-premise servers, and PC gamers repurposing their gaming rigs for inference. Hardware researchers can also use the tier lists to compare how different GPUs handle the same model, though the comparison tool is basic and best suited for quick side-by-side checks rather than deep analysis. For upgraders, CanIRun.ai helps answer a concrete question: if I buy an RTX 4090, which models that currently fail on my GTX 1080 will now run, and at what quantization? This is more actionable than generic VRAM calculators because it uses your actual detected hardware as the baseline.

Limitations are worth noting. The accuracy of hardware detection depends entirely on the browser's WebGPU implementation, which varies across browsers and operating systems. On some older GPUs or less common configurations, the tool may misreport VRAM or fail to detect the GPU entirely. It also does not test inference speed, token generation rates, or thermal behavior under load. A model that fits in VRAM may still run too slowly to be useful on an older GPU. Additionally, the database, while broad, is not exhaustive. Custom models, fine-tuned variants, or very recent releases may not appear, and there is no way to upload a custom model for analysis. The tool is free with no apparent paid tier, which is a plus, but also means support and update frequency are opaque.

In practice, CanIRun.ai works best as a first-pass filter. Use it to narrow down which models and quantization levels are worth downloading, then test actual performance with a tool like Ollama or LM Studio. It does not replace hands-on benchmarking, but it saves you from wasting bandwidth and disk space on models that your hardware cannot run at all. For anyone serious about local AI inference, it is a pragmatic, no-fuss companion that answers one question well: can I run it?