LLM quantization

Reddit 用户把大模型压进 60 MB 跑 CPU — 数字漂亮，我们先打个问号

SHADOW-250M: 250M params in 60 MB, 400 tokens/s on CPU. If reproducible, local AI costs drop fast — but single source, no benchmarks yet.