Users pick models/quantizations too large for their machine and hit thermal throttling and hardware-damage anxiety with no safe-operating guidance

Summary

A user running a 14B model on an i5-13450HX / RTX 4050 laptop sees cores pinned at 100% and CPU temperatures of 95°C, and has no idea whether this is normal, whether sustained inference is safe, or whether it is degrading the hardware. The root cause is exactly the gap this topic targets: without a VRAM/spec-aware recommender, users pick a parameter count and quantization that silently spills to CPU (or offloads poorly), turning a 'fits' model into a sustained thermal load. There is no tool that says 'with 6GB VRAM and this CPU, 14B at Q4 will offload heavily and throttle — run 7B–9B at Q5 instead, with these power limits.' Forum advice is contradictory and machine-specific, and Ollama/llama.cpp give no fit warning, no estimated tokens/sec, and no thermal or safety guidance. The pain is high-intensity because it combines fear of destroying an expensive laptop with an unusable, crawling inference session, and users would pay for a confident 'safe and optimal for your hardware' answer before spending hours downloading.