gbuzhf/Ornith-1.5-35B-A3B-Abliterated-MTP-UD-APEX-GGUF

Ornith-1.5-35B-A3B Abliterated — MTP + UD + ICE + APEX GGUF

Nine Pareto-optimal tiers of the abliterated model, each with an MTP head grafted in from the original Ornith-1.5 (the abliterated source ships none).

Measured on wikitext-2-raw, 16 chunks x 2048 ctx, against two references, so quantization damage and abliteration damage can be told apart.

How much did abliteration itself change the model?

Mean KLD (abliterated BF16 vs original BF16)	0.0151
same top-1 token	95.04%

For scale, the best quantization in this ladder costs ~0.022 KLD. Abliteration is a smaller perturbation than Q6_K quantization.

Tiers (9/9 measured)

tier	size	mean KLD	99% KLD	99.9% KLD	PPL ratio	same top-1	active bpw	file bpw	overall
MTP-UD-Q6_K	30.21 GB	0.0222	0.226	0.695	0.9960	94.05%	8.063	6.804	96.7
MTP-UD-Q5_K_S	25.84 GB	0.0261	0.269	0.899	0.9871	93.51%	7.693	5.820	96.3
TIEL_Calibrated-25G-ICE	24.85 GB	0.0276	0.269	1.061	0.9847	93.33%	7.686	5.597	96.1
MTP-25G-ICE	24.85 GB	0.0293 0.0281	0.300 0.279	1.280 1.097	0.9856 0.9910	93.44% 93.40%	7.686	5.597	96.0 96.1
TIEL_Calibrated-23G-ICE	22.84 GB	0.0321	0.326	1.173	0.9907	92.89%	7.523	5.143	95.7
MTP-23G-ICE	22.84 GB	0.0345 0.0329	0.345 0.339	1.151 1.240	0.9902 0.9941	92.39% 92.83%	7.523	5.143	95.4 95.6
TIEL_Calibrated-21G-ICE	20.85 GB	0.0392	0.421	1.606	0.9997	92.27%	7.357	4.695	95.0
MTP-21G-ICE	20.85 GB	0.0398 0.0395	0.400 0.412	1.655 1.726	0.9943 0.9929	92.06% 92.12%	7.357	4.695	94.9 95.0
TIEL_Calibrated-19G-ICE	18.82 GB	0.0576	0.602	2.123	1.0015	90.42%	7.192	4.240	93.3
MTP-19G-ICE	18.82 GB	0.0612 0.0585	0.630 0.598	2.327 1.955	1.0025 0.9995	90.02% 90.21%	7.192	4.240	93.0 93.2
MTP-UD-IQ4_XS	18.68 GB	0.0706	0.695	2.645	1.0502	89.40%	6.762	4.209	92.2
MTP-APEX-I-Compact-v2D-lite	17.57 GB	0.0925	0.893	3.141	1.0169	87.97%	5.228	3.956	90.5
MTP-APEX-I-Mini-v2D-lite	14.37 GB	0.2546	2.410	5.578	1.2129	80.70%	4.180	3.208	80.0

The four ICE rows were rebuilt on 2026-08-24 and the struck-through values are what they replaced. The new figures are measured at 64 chunks, the rest of the table at 16; a higher chunk count tightens the error bar without moving the mean. What changed, why, and the full measurement set: MEASUREMENTS-ICE-rebuild.md.

The four TIEL_Calibrated rows were added on 2026-08-30. Identical builds to their MTP-*-ICE counterparts — same abliterated trunk, same norm-fixed MTPv2 head, same ICE-base recipes, same byte budgets — but built with the importance matrix and embedded chat template from peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF, with credit to that repo for both. Its imatrix is 3000 x 512 = 1,536,000 tokens against the 573 x 512 = 293,376 behind the others. Also measured at 64 chunks.

Sorted best -> worst by overall (BF16 = 100), the same composite used on the non-abliterated card: 0.70/(1+meanKLD) + 0.30*sameTop1. All KLD columns are measured against the abliterated BF16, i.e. they isolate what the quantization costs.

Read the tail columns with care. 99.9% KLD is the ~33rd-worst token out of 32,768 — an extreme order statistic with large sampling variance, so it inverts between adjacent tiers without that meaning anything. 99% KLD rests on ~328 tokens and orders all nine tiers monotonically; mean KLD uses all 32,768 and separates the closest pair by 4.3 sigma. Rank on mean KLD; treat the tail columns as shape, not order.

Abliterated vs non-abliterated, same recipe

Both columns are the original 16-chunk measurements, so the delta isolates abliteration. The four ICE tiers have since been rebuilt; their current numbers are in the tier table above and in MEASUREMENTS-ICE-rebuild.md.

tier	KLD abl	KLD clean	Δ	top-1 abl	top-1 clean	Δ	PPL ratio abl	PPL ratio clean	Δ
MTP-UD-Q6_K	0.0222	0.0221	+0.6%	94.05%	93.85%	+0.20 pp	0.9960	0.9957	+0.0003
MTP-UD-Q5_K_S	0.0261	0.0272	-4.2%	93.51%	93.51%	+0.01 pp	0.9871	0.9862	+0.0009
MTP-25G-ICE	0.0293	0.0303	-3.4%	93.44%	93.16%	+0.28 pp	0.9856	0.9814	+0.0042
MTP-23G-ICE	0.0345	0.0361	-4.4%	92.39%	92.65%	-0.26 pp	0.9902	0.9885	+0.0018
MTP-21G-ICE	0.0398	0.0412	-3.4%	92.06%	92.03%	+0.03 pp	0.9943	0.9924	+0.0019
MTP-19G-ICE	0.0612	0.0608	+0.5%	90.02%	90.32%	-0.30 pp	1.0025	1.0030	-0.0005
MTP-UD-IQ4_XS	0.0706	0.0723	-2.4%	89.40%	89.46%	-0.06 pp	1.0502	1.0526	-0.0024
MTP-APEX-I-Compact-v2D-lite	0.0925	0.0954	-3.0%	87.97%	87.83%	+0.14 pp	1.0169	1.0101	+0.0068
MTP-APEX-I-Mini-v2D-lite	0.2546	0.2608	-2.4%	80.70%	80.49%	+0.21 pp	1.2129	1.2281	-0.0151

Δ is near zero or slightly negative across the ladder: abliteration does not make this model harder to quantize, and in the mid-range it is marginally easier — plausibly because projecting a direction out of ffn_down narrows its dynamic range.

Two further results, measured against the original BF16 as well (full numbers in KLD_RESULTS.txt):

• The two damages are strongly sub-additive — abliteration and quantization are largely orthogonal, so the combined figure sits far below their sum.
• No tier un-abliterates. Across the 9 rungs measured, the distance to the original BF16 stays above abliteration's own distance (0.0151), so quantization never pulls the model back toward the refusal behaviour.

Which tier is which

family	what it is
UD-*	Unsloth Dynamic 2.0 maps, replayed 1:1. Pins attention, the shared expert and token_embd at Q8_0 at every size and moves only the routed experts.
ICE-*	Bits allocated by how far a quantization error travels, not by activation magnitude. Named by target size.
APEX-I-*-v2D-lite	mudler's APEX maps plus one extra step on attn_k/attn_v in the ten full-attention blocks and on the output head.

Rule of thumb: Q6_K / Q5_K_S near-lossless, 25G/23G-ICE the quality sweet spot, 21G/19G-ICE the best small tiers, Compact/Mini only if you are tight on VRAM — Mini drops sharply.

The ICE tier, and why this is a 9-tier release

ICE allocates bits by error travel distance: how far a quantization error propagates before it reaches the output. Tensors writing straight into the residual stream, the always-on dense path, and the router are protected; the routed expert stack — 93% of the parameters but only 8-of-256 active per token — is left uniform. Routers stay F32, and the draft block is un-pinned so its experts follow the tier.

On the non-abliterated ladder, measured on this model and this harness, ICE lands ahead at matched size:

comparison	result
23G-ICE vs UD-Q4_K_XL (same size)	-5.0% KLD
23G-ICE vs APEX-I-Quality	-13.0% KLD and 0.61 GB smaller
25G-ICE vs APEX-I-Balanced	-12.2% KLD and 1.15 GB smaller

UD-Q4_K_XL,APEX-I-Quality-v2D-lite and APEX-I-Balanced-v2D-lite are each already covered by an ICE tier that is both smaller and closer to BF16, so rebuilding them would add size without adding a quality point. Full derivation, the refuted ffn_down rule and the measured convexity bound are in the original Ornith-1.5 card.

The 9-Tier Standard

From this release onward these nine recipes are the standard ladder: UD-Q6_K · UD-Q5_K_S · 25G-ICE · 23G-ICE · 21G-ICE · 19G-ICE · UD-IQ4_XS · APEX-I-Compact-v2D-lite · APEX-I-Mini-v2D-lite

They are the measured Pareto frontier of a 12-tier sweep on this architecture: every dropped tier is beaten on both size and KLD by one that ships. Reference measurements and methodology: Ornith-1.5-35B-MTP-UD-APEX-GGUF.

What was done to the source

PocketAiHub/Ornith-1.5-35B-A3B-Abliterated-GGUF is faithful at the tensor level — every tensor except ffn_down is byte-identical to ornith-ai's BF16. Abliteration is confined to ffn_down (both routed and shared experts), layers 15-39, at ~1.6-1.9e-02 L1-relative. Routers, attention, ffn_gate, ffn_up and layers 0-14 are untouched.

Three things were repaired while grafting:

1. MTP head restored — 20 blk.40.* tensors from ornith-ai's BF16; block_count 40 -> 41, nextn_predict_layers added.
2. tokenizer.ggml.add_bos_token restored to False — the source omits the key entirely, so llama.cpp falls back to its own default and tokenises differently from the original. Left unfixed this also invalidates any KLD against the original.
3. tokenizer.chat_template restored — the source ships a 7536-byte copy with the multi-system-message merge block removed; the original is 7828 bytes.

imatrix

bartowski's Ornith-1.5-35B-A3B-imatrix.gguf, reused unmodified. Justified by measurement, not assumption: the routers are byte-identical between the original and the abliterated model, so the same experts fire and the per-channel statistics still apply.

Recipes

All nine tensor maps were confirmed byte-exact against the corresponding shipped non-abliterated tier before this build, so the two ladders are directly comparable:

• UD — replayed 1:1 from unsloth/Ornith-1.0-35B-GGUF (Unsloth Dynamic 2.0).
• APEX v2D-lite — mudler's Ornith-1.5 maps, with attn_k/attn_v on the ten full-attention blocks and the output head each raised one step.
• ICE — bits allocated by how far a quantization error travels rather than by activation magnitude; the expert stack is uniform and the draft block un-pinned.

MTP / speculative decoding

Every tier carries the head, pinned Q8_0 (F16 attn_k/attn_v on the ICE tiers). It is grafted from the original model, so the draft head is not abliterated while the trunk is — worth knowing if you rely on the refusal behaviour under drafting: the head only proposes, the abliterated trunk verifies, so accepted tokens are always the trunk's.

Measured draft acceptance: 96.77% (390/403) on 23G-ICE, --spec-type draft-mtp, text-only.

prompt set	acceptance
code-novel	98.45%
structured	98.29%
copy-edit	97.30%
prose-novel	91.57%

Raw run in gate_spec_bench.json. Acceptance depends on the prompt mix — compare only against numbers taken on the same harness.

llama-server -m Ornith-1.5-35B-A3B-Abliterated-MTPv2-21G-ICE.gguf \
  --mmproj mmproj-Ornith-1.5-35B-A3B-Abliterated-F16.gguf \
  -c 8192 -fa on --jinja \
  --spec-type draft-mtp --spec-draft-n-max 1 --spec-draft-n-min 0 --spec-draft-p-min 0.75

MTPv2 — updated draft head (2026-08-23)

ornith-ai uploaded a newly trained native MTP head to the official Ornith-1.5-35B-A3B repo, so this repo is updated with it.

The nine MTPv2-* files are the same nine recipes as the previous ladder — same abliterated trunk, same imatrix, same per-tensor rule files, same pinned llama.cpp build. Only blk.40 is re-grafted. Every tensor outside blk.40, and the whole metadata block, is byte-identical to its MTPv1/ counterpart — checked tensor by tensor before anything was uploaded.

The four ICE tiers were later rebuilt again, on 2026-08-24, with a revised expert placement and a corrected blk.40 norm conversion. Their rows in the tier table show both the old and the new numbers.

What actually changed. The head in the original checkpoint had initializer-like weights; the replacement is trained. Read straight off the two published files:

blk.40 tensor	MTPv1	MTPv2
nextn.shared_head_norm.weight	1.02281 ± 0.00017	2.92531 ± 0.29643
attn_norm.weight	1.00006 ± 0.00302	0.90492 ± 0.15616
nextn.hnorm.weight	1.01571 ± 0.00558	0.49370 ± 0.09479
attn_q_norm.weight	1.00390 ± 0.00680	1.76704 ± 0.33864

The previous ladder is unchanged and still here, under MTPv1/ — same files, same digests, nothing deleted.

Standalone draft head

The new head is also published on its own, for use as a --model-draft sidecar instead of an embedded head:

file	size
mtp-Ornith-1.5-35B-A3B-Abliterated-MTPv2-BF16.gguf	3.74 GB
mtp-Ornith-1.5-35B-A3B-Abliterated-MTPv2-Q8_0.gguf	1.99 GB
mtp-Ornith-1.5-35B-A3B-Abliterated-MTPv2-Q4_K_M.gguf	1.26 GB

The BF16 one is a straight extraction from the same grafted master the ladder was quantized from, so its head is bit-identical to the head inside the MTPv2-* files.

Vision

Not re-hosted — use the projector from the source repo: mmproj-Ornith-1.5-35B-A3B-Abliterated-F16.gguf (0.90 GB). Without it the model is blind. Note --mmproj force-disables ctx_shift and cache_reuse.

Also included

sha256sums-MTPv2.txt and MANIFEST-MTPv2.txt for the current ladder, MEASUREMENTS-ICE-rebuild.md for the 2026-08-24 ICE rebuild, and the MTPv1 originals under MTPv1/ (KLD_RESULTS.txt, gate_spec_bench.json, sha256sums.txt, MANIFEST.txt).

The BF16 masters are not re-hosted: the abliterated source is at PocketAiHub and the original at ornith-ai.

Credit: PocketAiHub for the abliteration, bartowski for the imatrix and for publishing its corpus, mudler for the APEX reference maps, Unsloth for the UD 2.0 maps, Shisa-AI for Final 12K KL distill MTP Head (ShisaMTP variants), Jzinno for Dflash2 sidecar, peculiar-ragdoll for Tiel iMatrix Calibration and Sharp Chat Template.