The mmap layer streaming approach is smart for working around memory limits. In practice though, 1.58-bit ternary quantization tends to degrade quality noticeably on reasoning-heavy tasks compared to 4-bit — curious if you've measured perplexity deltas at the 140B scale.