ModelLens maps model VRAM, hardware needs

This is the kind of niche tool local-LLM users keep rebuilding in spreadsheets, so even a rough version can save real time if the estimates are trustworthy.

• –The core value is practical: matching model size, quantization, and available VRAM is one of the most annoying parts of local inference planning.
• –Family-specific calculators are more useful than generic parameter-count rules because memory behavior varies a lot across architectures and runtimes.
• –The unfinished hardware-discovery feature could become the stronger moat if it turns into a searchable compatibility database instead of just another estimator.
• –The main risk is confidence: VRAM math is easy to overpromise, so the product will need clear assumptions, caveats, and maybe error bands.
• –For local AI builders, this sits squarely in the “small tool, high daily usefulness” category.