CanIRun.ai is a browser-based tool that uses WebGPU APIs to detect a user's hardware and instantly determine which open-weight AI models they can run locally — no installation, no registration, and no data leaving the device. Built by Miguel Angel Duran Garcia, known online as midudev, the site cross-references detected GPU and VRAM specs against a database of models from over a dozen AI organizations, assigning each a performance grade from S through F. The catalog spans the full range of local inference, from sub-1GB edge models like Qwen 3.5 0.8B and Llama 3.2 1B to multi-hundred-gigabyte Mixture-of-Experts (MoE) giants including DeepSeek R1 (671B), DeepSeek V3.2 (685B), and Moonshot AI's Kimi K2 (1 trillion parameters). Data is aggregated from the three dominant local inference ecosystems: llama.cpp, Ollama, and LM Studio.
Each model entry surfaces quantization options (Q2_K through F16), context window lengths, architecture type (Dense vs. MoE), and VRAM requirements at each precision level, giving users a practical breakdown of the tradeoffs between memory usage and model quality. Hacker News discussion flagged one area where the site's estimates may need refinement: MoE models like OpenAI's GPT-OSS 20B require full VRAM to load, but only activate a fraction of parameters per token — roughly 3.6B active out of 20B total — meaning generation speed is closer to a 3-4B dense model. Commenter meatmanek noted the site correctly reflects VRAM requirements but performance grades may not fully capture this speed advantage. Separately, practitioners such as mark_l_watson highlighted that small local models, particularly Qwen 3.5 9B, are well-suited for embedded tasks, tool use, and information extraction, while larger workloads like coding assistance remain more practical via cloud APIs.
Duran Garcia is a Google Developer Expert, Microsoft MVP, and GitHub Star with over 500,000 YouTube subscribers and 178,500 Twitch followers, making him one of the most-followed Spanish-language developer educators globally. His annual conference, miduConf, drew over 500,000 live viewers in its most recent edition. The tool launched free, with no monetization mechanism; his announcement described it as built "for the local AI community." CanIRun.ai runs entirely client-side, scores models per-quantization across roughly 40 GPU SKUs and 12 Apple Silicon chips, and covers models from 16 AI organizations. Duran Garcia has not announced a paid tier or commercial roadmap.