Huihui-Ornith-1.0-35B-abliterated — MTP + UD + APEX-v2D-lite GGUF

huihui-ai's abliterated Ornith-1.0-35B with an MTP head grafted in, the vision projector, and eight tiers.

Measured MTP draft acceptance: 86.4% (748/866), 44.1 t/s, --mmproj loaded.

This build is deliberately a one-variable change from the faithful gbuzhf/Ornith-1.0-35B-MTP-UD-APEX-GGUF: identical MTP head, identical tensor maps, identical imatrix. Only the trunk is abliterated. That is what makes the two comparable at all.

	faithful	abliterated	delta
MTP draft acceptance	87.4%	86.4%	−1.0 pp
tokens/s (gate)	49.1	44.1	−5.0

Same harness, same prompt mix, so this comparison is like-for-like. Abliteration cost about one point of draft acceptance — the un-abliterated head still fits the abliterated trunk nearly as well.

Tiers

Ranked by active bpw — routed experts weighted 8/256 because that is what actually fires per token, token_embd excluded (it is a get_rows, not a matmul). It is a measure of where the bits go, not a measured quality score. Every file size matched its pre-build prediction to 0.01 GB.

tier	GB	active bpw	file bpw
`MTP-UD-Q6_K`	30.21	8.063	6.804
`MTP-UD-Q5_K_S`	25.84	7.693	5.820
`MTP-UD-Q4_K_XL`	23.22	7.471	5.230
`MTP-APEX-I-Balanced-v2D-lite`	26.30	6.913	5.922
`MTP-UD-IQ4_XS`	18.68	6.762	4.209
`MTP-APEX-I-Quality-v2D-lite`	23.85	6.699	5.371
`MTP-APEX-I-Compact-v2D-lite`	17.57	5.228	3.956
`MTP-APEX-I-Mini-v2D-lite`	14.37	4.180	3.208

The MTP head is pinned Q8_0 in every tier (8.515 bpw measured). No imatrix can reach blk.40 — it is grafted after calibration and never runs during an imatrix pass — so it is unguided RTN either way, and draft acceptance converts directly into tokens/sec.

Note the two inversions against file size. UD-IQ4_XS (18.68 GB) scores above APEX-I-Quality-v2D-lite (23.85 GB) on the active path, and UD-Q4_K_XL (23.22 GB) above APEX-I-Balanced-v2D-lite (26.30 GB). The UD maps pin attention, shared experts and token_embd at Q8_0 at every tier, and roughly 55% of active parameters are non-routed — attention, the shared expert and the 508 M-param output head all run on every token, while routed experts contribute 8/256. If you are choosing on quality per GB, start from the top of this table, not from file size.

How the MTP head got here

huihui-ai publishes the abliterated model as bf16 GGUF only — there are no abliterated safetensors on the Hub. The usual path grafts 19 mtp.* tensors into safetensors and then converts, which was impossible, so the graft was done one stage later, directly at GGUF level:

+ 20 blk.40.* tensors, copied byte-for-byte from the faithful build's bf16 master
~ qwen35moe.block_count            40 -> 41
+ qwen35moe.nextn_predict_layers   (absent) -> 1

The donor head is Qwen3.6-35B-A3B's original MTP head (844.6 M params), which passed a byte-identity HEAD GATE in the faithful build — sha256 faac91f15cbe54475faa2578bedc46a7c29a947b8a3e7ef3ecd376ae079826ab. So this repo provably carries the same head as the faithful one.

Verified before building: huihui's 733 tensor names are identical to the reference, and all shapes agree once trailing 1s are normalised (huihui writes ffn_gate_inp_shexp as [2048, 1], convert_hf_to_gguf.py writes [2048] — the same 2048 values). A grafted-master gate then asserted 753 tensors, 20 in blk.40, block_count=41, nextn_predict_layers=1.

Vision

Huihui-Ornith-1.0-35B-abliterated-mmproj-F16.gguf (0.90 GB) is required — without it you have a blind model. It is huihui-ai's own projector, kept separate so one copy serves all eight tiers.

llama-server -m Huihui-Ornith-1.0-35B-abliterated-MTP-UD-Q5_K_S.gguf \
  --mmproj Huihui-Ornith-1.0-35B-abliterated-mmproj-F16.gguf \
  -c 8192 -fa on --jinja \
  --spec-type draft-mtp,ngram-mod \
  --spec-draft-n-max 1 --spec-draft-n-min 0 --spec-draft-p-min 0.75 \
  --spec-ngram-mod-n-min 8 --spec-ngram-mod-n-max 24 --spec-ngram-mod-n-match 48

--mmproj silently force-disables ctx_shift and cache_reuse.

Honest limits — read these

The imatrix was computed on the UN-abliterated model. It is the same 2-way Ornith-native merge (bartowski calibration_datav5 + unsloth, 2463 chunks, 1.26 M tokens) used by the faithful build. Abliteration changes activation statistics, so a base imatrix will misallocate precision to some degree, and on a 256-expert MoE it can misdescribe routing too. This was a deliberate trade to keep the build on a CPU box and to hold every variable except the trunk constant. It is the one known-suboptimal ingredient here.
Acceptance is one gate run, 866 draft tokens, on a code-novel + copy-edit prompt mix. Comparable to the faithful build's 87.4% because it is the identical harness — not comparable to acceptance figures quoted from other harnesses.
Abliteration is not free. It moves weights off the trained optimum by design, and that cost is separate from and additional to quantization error. If you do not need refusal removal, use the faithful build.

Also included