HotPin: Lossless 120B MoE Inference on 24GB RAM with 50 Lines of CPU Code?

A Hacker News submission introduces HotPin, with a title claiming “lossless” inference for a 120B-parameter Mixture-of-Experts model using only 24GB of RAM and roughly 50 lines of CPU-oriented code. The submission had a score of 14 and no comments at the supplied publication time. The available abstract does not identify the model, explain the memory strategy, report tokens-per-second, or provide benchmarks and implementation details. The headline is technically interesting, but the central claims remain unverified from the supplied source alone.