@joshfpocock: Someone ran a 2.78 TRILLION parameter model on a single CPU in 8.24 GB of RAM. No GPU, no PyTorch. The whole engine is 176 KB of C. The trick: 93% of the model is experts that never load. They stay on the SSD and only get read when a token routes to them. 5.5 TB at full precision → 8.24 GB resident, byte-identical output. The fine print: 33 seconds per token. Not 33 tokens per second. You also need a 1.56 TB SSD and it’s Linux only. One YouTuber built it and admitted on camera he never generated a single token — not enough disk space. Not a product. A proof. The memory wall was never about model size, it’s about which bytes you keep. #ai #opensource #llm #localai #cprogramming