No reliable VRAM-aware way to know if a model + quantization will actually run locally

Summary

Local LLM users with common 24GB GPUs are overwhelmed by new open-weight releases that are marketed as 'local' but require multi-GPU rigs costing tens of thousands of dollars. Model cards usually list total parameters, active parameters, and max context, but not the real memory footprint after choosing GGUF/GPTQ/AWQ quantization, accounting for KV-cache at the desired context length, runtime overhead, and vision projectors. Users must manually download, convert, try, and hit OOM errors before discovering a model does not fit. Existing hubs rank by popularity or raw size, not by whether a specific hardware + quantization + context configuration will actually work. This is a high-friction, high-cost guessing game that a VRAM-aware recommender could directly solve.