---

license: mit

base_model:

- huihui-ai/Huihui-Ornith-1.5-35B-A3B-abliterated

- peculiar-ragdoll/Cyber-Tiel-Coder-35B-A3B-GGUF-MTP

base_model_relation: quantized

pipeline_tag: text-generation

library_name: gguf

language:

- en

- zh

tags:

- gguf

- llama.cpp

- qwen35moe

- moe

- imatrix

- ICE

- mtp

- abliterated

- uncensored

---

Ornith-1.5-35B-A3B Abliterated — CyberTiel-Calibrated MTPv2 ICE

Four ICE tiers of huihui-ai/Huihui-Ornith-1.5-35B-A3B-abliterated,

quantized against a trunk-matched importance matrix, with the model's own trained

MTPv2 head and the Qwen-Sharp v22.4.1 chat template embedded.

The BF16 master they were all built from is published here too.

> ⚠️ Uncensored model. This is an abliterated checkpoint: refusal behaviour has been

> suppressed in the weights. It will comply with requests a stock model declines. You are

> responsible for how you use it. Sandbox it at the OS level and control its network and

> code-execution access — with no refusal backstop, a prompt injection from a hostile page

> or third-party code has nothing to stop it. If you want the guardrailed equivalent, use

> the non-abliterated TIEL-Calibrated ICE tiers.

Which one?

Sizes are picked to land between the familiar UD rungs, so use those as your anchor:

| file | size | fits (RAM+VRAM) | where it sits |

|---|--:|:--|---|

| …-19G-ICE.gguf | 18.82 GB | 24 GB | between UD-IQ4_XS and UD-Q4_K_S — most context headroom here |

| …-21G-ICE.gguf | 20.85 GB | 24 GB | UD-Q4_K_S class; the rung below Q4_K_M |

| …-23G-ICE.gguf | 22.84 GB | 24–32 GB | start here — UD-Q4_K_XL class, snug on 24 GB |

| …-25G-ICE.gguf | 24.85 GB | 32 GB | between UD-Q4_K_XL and UD-Q5_K_XL; closest to BF16 here |

| …-BF16.gguf | 71.07 GB | — | the master — bake your own tiers from it |

Rule of thumb: take the largest tier that leaves you the context you actually need at

q8_0 KV. If you already run a UD build of this model, pick the ICE tier at the same

size and keep everything else the same.

Measurements

KL divergence against the BF16 master published in this repo — one binary, one

reference, one session per table, 64 chunks at n_ctx 2048. UD-* rows are from

peculiar-ragdoll/Cyber-Tiel-Coder-35B-A3B-GGUF-MTP,

whose chat template this build also uses. Both ladders quantize the same

huihui-abliterated weights. The importance matrices differ: those tiers use his 800-chunk

matrix, these use his 27,513-chunk full-corpus one. Compared directly from the two files

they allocate bits almost identically (expert-tensor in_sum2 cosine 0.99569, top-1 %

channel overlap 0.948), so what the tables show is a recipe difference, not a calibration

one. Grouped by size band, largest file first, so the two tables line up row for row.

Bold marks the best file in each band.

English text — WikiText-2  ·  PPL(base) 7.577170

| file | size | mean KLD | 99% KLD | 99.9% KLD | PPL ratio | same top-1 | active bpw | file bpw | overall |

|---|--:|--:|--:|--:|--:|--:|--:|--:|--:|

| ≈ 25–27 GB | | | | | | | | | |

| UD-Q5_K_XL | 26.98 GB | 0.024029 | 0.2520 | 0.9873 | 0.9927 | 94.00 % | 7.833 | 6.077 | 96.6 |

| 25G-ICE | 24.85 GB | 0.027866 | 0.2793 | 1.0757 | 0.9867 | 93.36 % | 7.686 | 5.599 | 96.1 |

| ≈ 22.5–23 GB | | | | | | | | | |

| 23G-ICE | 22.84 GB | 0.033146 | 0.3304 | 1.1405 | 0.9864 | 92.68 % | 7.523 | 5.146 | 95.6 |

| UD-Q4_K_XL | 22.75 GB | 0.035699 | 0.3651 | 1.3015 | 0.9770 | 92.63 % | 7.474 | 5.123 | 95.4 |

| UD-Q4_K_M | 22.52 GB | 0.036217 | 0.3761 | 1.2884 | 0.9753 | 92.40 % | 7.130 | 5.073 | 95.3 |

| ≈ 21 GB | | | | | | | | | |

| UD-Q4_K_S | 21.28 GB | 0.040310 | 0.4156 | 1.5351 | 0.9795 | 91.98 % | 7.025 | 4.793 | 94.9 |

| 21G-ICE | 20.85 GB | 0.039589 | 0.4162 | 1.5265 | 0.9840 | 92.22 % | 7.357 | 4.698 | 95.0 |

| ≈ 18 GB | | | | | | | | | |

| 19G-ICE | 18.82 GB | 0.059336 | 0.6120 | 1.9325 | 0.9987 | 90.09 % | 7.192 | 4.241 | 93.1 |

| UD-IQ4_XS | 18.12 GB | 0.068901 | 0.7185 | 2.3927 | 1.0435 | 89.59 % | 6.757 | 4.082 | 92.4 |

Code — code.test.raw  ·  PPL(base) 2.194208

| file | size | mean KLD | 99% KLD | 99.9% KLD | PPL ratio | same top-1 | active bpw | file bpw | overall |

|---|--:|--:|--:|--:|--:|--:|--:|--:|--:|

| ≈ 25–27 GB | | | | | | | | | |

| UD-Q5_K_XL | 26.98 GB | 0.011993 | 0.1702 | 0.8170 | 1.0009 | 97.25 % | 7.833 | 6.077 | 98.3 |

| 25G-ICE | 24.85 GB | 0.015497 | 0.2207 | 1.0820 | 1.0007 | 97.06 % | 7.686 | 5.599 | 98.0 |

| ≈ 22.5–23 GB | | | | | | | | | |

| 23G-ICE | 22.84 GB | 0.019269 | 0.2822 | 1.3248 | 1.0018 | 96.66 % | 7.523 | 5.146 | 97.7 |

| UD-Q4_K_XL | 22.75 GB | 0.019742 | 0.2909 | 1.3855 | 1.0041 | 96.61 % | 7.474 | 5.123 | 97.6 |

| UD-Q4_K_M | 22.52 GB | 0.020440 | 0.2880 | 1.3215 | 1.0040 | 96.59 % | 7.130 | 5.073 | 97.6 |

| ≈ 21 GB | | | | | | | | | |

| UD-Q4_K_S | 21.28 GB | 0.023484 | 0.3466 | 1.6264 | 1.0028 | 96.31 % | 7.025 | 4.793 | 97.3 |

| 21G-ICE | 20.85 GB | 0.024669 | 0.3787 | 1.6756 | 1.0044 | 96.28 % | 7.357 | 4.698 | 97.2 |

| ≈ 18 GB | | | | | | | | | |

| 19G-ICE | 18.82 GB | 0.036910 | 0.5798 | 2.3285 | 1.0154 | 95.24 % | 7.192 | 4.241 | 96.1 |

| UD-IQ4_XS | 18.12 GB | 0.044161 | 0.6689 | 2.8208 | 1.0191 | 94.80 % | 6.757 | 4.082 | 95.5 |

overall = 0.70/(1+meanKLD) + 0.30*sameTop1. Sizes are measured, not predicted.

> Don't compare the two tables to each other. Code is more predictable text, so every

> file scores about half as much divergence on it. Compare rows within a table.

Picking a tier

!ICE vs Unsloth Dynamic — file size against quantization loss, on both corpora

Both ladders are cut from the same BF16 master, so the only question is which one loses less

of it at a given file size.

| tier | size | English text | code |

|---|--:|:--|:--|

| 19G-ICE | 18.82 GB | saves 0.18 GB | saves 0.20 GB |

| 21G-ICE | 20.85 GB | saves 0.64 GB | saves 0.18 GB |

| 23G-ICE | 22.84 GB | saves 0.70 GB | saves 0.12 GB |

| 25G-ICE | 24.85 GB | saves 0.55 GB | level |

saves X GB = a UD file has to be that much larger to lose as little as this tier does.

On code the two ladders run close together above 22 GB.

UD-Q5_K_XL (26.98 GB) is the most faithful file measured here. If you have 27 GB to

spend, take it — ICE is aimed at the 18–25 GB band, where it buys you the gap between two UD

rungs without the extra gigabyte.

Corpora

| corpus | size | contents |

|---|--:|---|

| wiki.test.raw | 1.29 MB | WikiText-2 raw test, as distributed with the llama.cpp CI dataset |

| code.test.raw | 1.09 MB | 70 source files from 5 repos pinned to commit SHAs — Rust 42.5 % (ripgrep, bat), C/C++ 21.9 % (llama.cpp), Python 20.8 % (requests), Go 14.8 % (gin) — all MIT / Unlicense / Apache-2.0 |

code.test.raw is in corpora/ with the script that rebuilds it

byte-exactly; no dataset gating, no account needed. Raw llama-perplexity output for every

row of both tables is under measurements/.

Two bpw columns. Only 8 of 256 experts fire per token, so a bit in ffn_*_exps is worth

about 3 % of a bit in attention, the shared expert or the output head — those run on every

token. active bpw weights by that (routed experts at 8/256, token_embd excluded, MTP

scored separately); file bpw is just size ÷ parameters. It is why a 22.84 GB file computes

at ~7.5 bpw, and it is the more useful column when comparing two files of similar size.

> KLD measures fidelity to this repo's master and nothing else — not reasoning, tool use or

> speed. It is also within-lineage: every ladder is measured against its own master, so

> these values are not comparable to another repo's. Compare across lineages with PPL.

What ICE is, briefly

ICE = Isolation of Compounding Error: allocate bits by **how far a quantization error

travels through the network**, rather than by activation magnitude. Two visible consequences:

• A uniform expert stack — ffn_gate, ffn_up and ffn_down share a type at each

depth. Controlled tests at identical size favoured uniform here: ffn_down writes

straight into the residual stream, while gate/up errors are attenuated through the

gating nonlinearity — and there is twice as much gate+up mass paying for any bump.

• The draft block follows the tier — blk.40's experts are not pinned high; only its

projections stay Q8_0, which frees ~0.4 GB with draft acceptance intact.

Method, measurements and the cases where it does not win:

gbuzhf/ICE-quantization (Apache-2.0).

What's inside

• MTPv2 head, native. The abliterated checkpoint already carries

Ornith-1.5's trained head: its mtp.* norms are bit-identical to ornith-ai's on all seven

tensors (means 0.2734–2.9251; an untrained head sits at 1.000 ± 0.003). Every tier ships

blk.40 with nextn.* intact for speculative decoding.

• Importance matrix: trunk-matched. Cyber-Tiel-Coder-35B-A3B-fullcorpus.imatrix.gguf

— 27,513 chunks × 512 (~14.1 M token-passes), measured by peculiar-ragdoll on this exact

abliterated trunk. Weighted toward security and code, so expect it to favour that work.

• Chat template: Qwen-Sharp v22.4.1, embedded (29,674 B), add_bos_token=false.
• Vision works — the projector just isn't duplicated here. These tiers see images once

you load Ornith-1.5's own projector alongside them. Grab

mmproj-Ornith-1.5-35B-BF16.gguf

from ornith-ai/Ornith-1.5-35B-A3B-GGUF

and pass --mmproj. It is one file shared by every tier, unaffected by the abliteration

and by quantization — a 19G tier sees exactly as well as a 25G one.

Run it

hf download gbuzhf/Ornith-1.5-35B-A3B-Abliterated-CyberTiel-Calibrated-MTPv2-ICE-GGUF \
  Ornith-1.5-35B-A3B-Abliterated-CyberTiel_Calibrated-MTPv2-23G-ICE.gguf --local-dir .
llama-server -m Ornith-1.5-35B-A3B-Abliterated-CyberTiel_Calibrated-MTPv2-23G-ICE.gguf \
  -ngl 99 --jinja --ctx-size 262144 -fa on -ctk q8_0 -ctv q8_0 \
  --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0

If a tier does not fit your VRAM, do not lower -ngl. Keep every layer on the GPU and

push only the routed experts to CPU with --n-cpu-moe N; each +1 frees roughly 0.5 GB.

With vision — add Ornith-1.5's projector:

hf download ornith-ai/Ornith-1.5-35B-A3B-GGUF mmproj-Ornith-1.5-35B-BF16.gguf --local-dir .
llama-mtmd-cli -m Ornith-1.5-35B-A3B-Abliterated-CyberTiel_Calibrated-MTPv2-23G-ICE.gguf \
  --mmproj mmproj-Ornith-1.5-35B-BF16.gguf -ngl 99 \
  --image screenshot.png -p "What is failing here?"

--mmproj works with llama-server too, so vision is available over the API.

Credits

• ornith-ai — Ornith-1.5-35B-A3B, and the trained MTPv2 head these tiers carry
• huihui-ai — the abliteration this build quantizes
• peculiar-ragdoll — the full-corpus cyber/code importance matrix and the Qwen-Sharp chat template
• eaddario — calibration corpus components (MIT)
• ggml-org/llama.cpp — quantizer and measurement tooling
• Qwen team — the underlying architecture