Did anyone else notice the Ornith 1.5 35B GGUFs got a "silent" update?

Hey all, I stumbled onto something funny and figured I'd share before it gets buried. I'd been running the Ornith 1.5 35B A3B GGUF (Q4_K_M) for a while, then recently a new revision showed up in my HF cache and I figured, eh, I'd pruned my old one. Fast forward a bit later when I wanted to dig into what actually changed, and it turned into a little rabbit hole. Long story short: the Aug 24 "update" to the official GGUF repo (with no meaningful commit message other than the pretty uninformative "Upload with huggingface_hub") wasn't just a re-upload of the same files. I pulled my old snapshot out of a Btrfs backup and diffed it against the current one, and here's what's actually different in the Q4_K_M: The weights themselves changed. I compared the raw tensor bytes and they differ from the very first byte. So it's not the same quantization with a fresh coat of paint. The importance matrix / calibration is different. The GGUF metadata still has the author's machine paths baked in, and they changed from something like 35b-4000 to ornith-1.5-35b . Which lines up with the fact that the old file had general.version = "4000" and general.finetune = "35b" — fields the new one dropped. The labels got cleaned up. size_label went from "256x2.6B" to "35B" , and they added license , tags , and a basename . Just to preempt the usual comment: I verified this against the official ornith-ai repo specifically, not the third-party mirrors, so your copy might already differ if you pulled from somewhere else. The interesting bit: everything structural is identical — same architecture ( qwen35moe ), same 248k tokenizer, same file_type , same expert counts, same context length. So it's the same model family, just re-quantized against a different calibration checkpoint and relabeled. The whole file only shifted by a couple hundred bytes, which is exactly why nobody seems to have noticed. Anyway, I mostly want to know if anyone else saw this roll out, and whether it's noticeable in practice. The new calibration path ( ornith-1.5-35b ) makes me curious whether it's actually better or just tidier. Anyone run both? (Also shoutout to Btrfs snapshots for saving my ass here, hf prune is great but it does, uh, prune.) EDIT: I wonder if this update is about fixing the "random" MTP head from this post .