Withdrawn — Ornith-1.0-35B-A3B MTP + UD/APEX GGUF

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

What was here

Eight GGUF tiers of ornith-ai/Ornith-1.0-35B-A3B published 2026-08-12 — four Unsloth-Dynamic + MTP and four APEX-I-v2D-lite + MTP — together with the BF16 master, the vision projector, the MTP donor head and a merged importance matrix. The MTP block was pinned Q8_0 in every tier.

Why it was withdrawn

Superseded by the Ornith-1.5 line. Keeping a quarter-terabyte of a superseded ladder online was not a good use of Hugging Face's free public storage, which is offered on a best-effort basis for content of ongoing value to others.

Where to go instead

you want	go to
the same recipes on the current base	gbuzhf/Ornith-1.5-35B-MTP-UD-APEX-GGUF
the abliterated Ornith-1.0 line (still published)	gbuzhf/Huihui-Ornith-1.0-35B-abliterated-MTP-UD-APEX-GGUF
the upstream model	ornith-ai/Ornith-1.0-35B-A3B

Everything here was reproducible from the upstream weights with llama.cpp; nothing unique to this repository is lost beyond the convenience of the prebuilt files.

Two corrections that belonged to this release

Recorded here because they were published and should not quietly disappear:

Tensor arithmetic does not prove a base-model version. Qwen3.5-35B-A3B and Qwen3.6-35B-A3B share layer/expert/dimension counts and the Qwen3_5MoeForConditionalGeneration architecture string — a class name, not a version marker. Only a control comparison against the other candidate settles it.
Draft acceptance does not track lineage proximity. This was asserted and then refuted: a checkpoint closer to the shared base accepted lower. Acceptance figures are also harness-relative and must not be compared across harnesses.

Citation