Withdrawn — KAT-Coder-V2.5-Dev REAP-205E MTP GGUF

These artifacts were removed on 2026-09-12. The repository is kept only so the DOI continues to resolve.

What was here

An expert-pruned build of Kwaipilot/KAT-Coder-V2.5-Dev published 2026-08-10: 256 → 205 experts per layer (19.92% pruned) by REAP saliency, with an MTP head grafted, a BF16 master and nine quantized tiers.

Why it was withdrawn

A publication gate was skipped. The build procedure made draft-acceptance measurement a hard stop — publish only above ~0.80 — and that measurement was never run. No acceptance figure for this pruned model exists. The model card cited the unpruned model's 0.90–0.94, which reads as validation of something that was never tested. A separate open defect (wrong-language output) was also never closed.

Re-verification in 2026-08-15 confirmed the methodology was sound: the keep-index is bit-exact REAP top-205/layer across all 40 layers, every rate-table number reproduces, the GGUF structure passes all checks, and recipe parity against the unpruned release is exact — 753 tensors, zero name or quant-type differences. The pruning was done correctly; the resulting model was never shown to work. Those are different claims, and only the first one was ever supported.

Withdrawing is the honest resolution. The alternative — leaving weights online whose headline capability claim was borrowed from a different model — is worse than removing them.

Where to go instead

you want	go to
the unpruned model, same recipes, measured	gbuzhf/KAT-Coder-V2.5-Dev-MTP-GGUF
the upstream model	Kwaipilot/KAT-Coder-V2.5-Dev

If you want to reproduce the pruning

The method is reproducible without these files. The saliency measurement and keep-index derivation are documented, and the pruning itself is deterministic given the saliency tensor. Anyone repeating it should run the acceptance gate that this release did not.

Citation