With 256GB of dual-channel DDR5 behind it, how big an AI model can the 9900X actually run locally, and at what speed?

  The 9900X's real asset is its 256GB ceiling: dual-channel bandwidth cannot feed the biggest models at speed, but capacity can hold very large ones, so in pure-CPU inference the resident model tier follows the RAM installed, with 12 cores/24 threads at 5.6GHz and AVX-512 carrying prefill and token generation while data preprocessing runs alongside without dropping the ball.

  To genuinely eat 256GB you must accept bandwidth as the wall: tens-of-billions-class models load fine, yet generation speed is still set by memory bandwidth. It suits people who prioritize loading the model over tokens per second; anyone chasing per-token speed should buy a GPU. Size your model first; that number tells you if the 256GB earns its keep.