Same question every time a new model drops: can my GPU actually handle this? Stumbled on can i run ai, which compares your hardware (VRAM, RAM, CPU) against the requirements of popular open models — Llama, Mistral, Qwen, Stable Diffusion, etc. It also estimates rough tokens-per-second.

Curious what tooling other people lean on for this:

• Do you trust the VRAM estimates these calculators give, or have you found them too optimistic once you account for KV cache at longer context lengths?
• For folks running quantized models (Q4 / Q5 / GGUF), how much wiggle room do you actually get vs the FP16 baseline numbers most spec pages cite?
• Has anyone seriously benchmarked Apple Silicon (M-series) on 30B+ models in real workflows? Unified memory spec sheets always look great but reports of real throughput seem mixed.

Not affiliated, sharing in case it's useful — definitely cleaner than spreadsheet-checking VRAM tables by hand every time HuggingFace ships something new.