I built a small site for the local-model crowd: https://aimodelradar.dev — it tells you which open models actually fit your GPU or Mac, measured rather than guessed.
How it works: weights = the byte size of the real published GGUF file (linked next to every model — many of them are unsloth’s quants, thank you 
); context cache computed layer by layer from each model’s config.json (hybrid attention, sliding windows, MLA, DeepSeek sparse attention); runtime overhead plus a stated safety margin. If a config is gated or the architecture is one it can’t model, it says UNKNOWN instead of inventing a number. Macs are planned at 70% of unified memory because macOS caps what the GPU may take.
Also on it: a momentum board built from Hub trending/downloads/likes with a published formula (no benchmarks, on purpose), hourly GPU rental medians, a three-way cost calculator (API vs owned hardware vs rented GPU), a plain-language glossary in EN/ES/PT, a budget-hardware guide — and a read-only JSON API plus an MCP server you can plug into Claude/Cursor: Developers — JSON API and MCP server — AI Model Radar
Data: the Hub’s public model metadata, OpenRouter (CC BY 4.0) and Vast.ai, collected several times a day; every number on the site carries its collection time, and the methodology page (linked in the footer) lists what it refuses to publish.
It tracks 51 first-party models right now. If yours is missing, or a number looks wrong, tell me — a correction that survives checking updates the data. Which model should it look at next?