---

license: mit

base_model:

• ornith-ai/Ornith-1.5-35B-A3B

base_model_relation: quantized

pipeline_tag: text-generation

library_name: gguf

language:

• en
• zh

tags:

• gguf
• ICE
• ice-quant
• imatrix
• sangreal
• mtp
• speculative-decoding
• moe
• qwen35moe
• abliterated
• uncensored

---

Ornith-1.5-35B-A3B — Huihui × Sangreal ICE

Five ICE tiers of huihui-ai's abliterated Ornith-1.5-35B-A3B, quantized against

Sangreal — a purpose-built 77-bucket calibration corpus — and shipping

peculiar-ragdoll's Qwen-Sharp chat template. The abliterated checkpoint already

carries Ornith-1.5's trained MTPv2 head, so every tier drafts for speculative

decoding out of the box.

This release introduces a new ICE rung: 15G-ICE, the first tier below the

17 GB bound the method was originally scoped to, built for 16 GB cards. The BF16

master all five were cut from is published

here.

Best got better — ICE 1.6 (2026-09-25)

Rebuilt tiers vs the files they replace — mean KLD to BF16, paired on the same chunks, one session.

| tier | code KLD (2K) | code @16K | paired code t | |

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

| 15G | 0.088158 → 0.083796 (−4.9%) | −4.4% | −6.02 | clear win - replaced |

| 19G | 0.036887 → 0.034761 (−5.8%) | −6.7% | −3.07 | clear win - replaced |

| 23G | 0.018988 → 0.018473 (−2.7%) | −4.5% | −1.49 | replaced · lower KLD at 2K and 16K; code top-1 −0.06 pp |

| 25G | 0.015308 → 0.014603 (−4.6%) | −5.2% | −2.61 | clear win - replaced |

Changed: output.weight Q8_0 → Q6_K, MTP head at the floor, freed bytes → routed experts, 36,025-chunk Sangreal imatrix, template v22.5.1.

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

> suppressed in the weights. 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.

Measurements

One binary, one reference, one session, 64 chunks at n_ctx 2048. Every file below —

including our previous CyberTiel ladder and the Official CyberTiel's UD ladder — was

re-measured in this same session; cross-session KLD drifts ~0.8% on identical inputs,

which is the same order as the effect being measured.

2026-09-25: a second session (same GPU class, binary, reference, corpora) re-measured every shipped

Sangreal tier and every UD row: all 24 values reproduced exactly. Replaced rows show ~~then~~ now.

Raw logs: measurements/2026-09-25/.

Table 1 — Code (code.test.raw)

PPL(base) 2.194208 · BF16 = 100

| 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 | 6.080 | 98.3 |

| (New) Sangreal 25G-ICE | ~~24.85 GB~~<br>24.82 GB | ~~0.015308~~<br>0.014603 | ~~0.2274~~<br>0.2129 | ~~1.0218~~<br>0.9268 | ~~1.0017~~<br>1.0014 | ~~97.00 %~~<br>97.03 % | ~~7.686~~<br>7.385 | ~~5.599~~<br>5.593 | ~~98.0~~<br>98.1 |

| CyberTiel 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 | | | | | | | | | |

| (New) Sangreal 23G-ICE | ~~22.84 GB~~<br>22.81 GB | ~~0.018988~~<br>0.018473 | ~~0.2831~~<br>0.2767 | ~~1.2650~~<br>1.2717 | ~~1.0023~~<br>1.0026 | ~~96.73 %~~<br>96.67 % | ~~7.523~~<br>7.215 | ~~5.145~~<br>5.139 | 97.7 |

| CyberTiel 23G-ICE | 22.84 GB | 0.019269 | 0.2822 | 1.3248 | 1.0018 | 96.66 % | 7.523 | 5.145 | 97.7 |

| UD-Q4_K_XL | 22.75 GB | 0.019742 | 0.2909 | 1.3855 | 1.0041 | 96.61 % | 7.474 | 5.126 | 97.6 |

| UD-Q4_K_M | 22.52 GB | 0.020440 | 0.2880 | 1.3215 | 1.0040 | 96.59 % | 7.130 | 5.075 | 97.6 |

| ≈ 21 GB | | | | | | | | | |

| Sangreal 21G-ICE | 20.85 GB | 0.023415 | 0.3529 | 1.4214 | 1.0042 | 96.35 % | 7.357 | 4.698 | 97.3 |

| UD-Q4_K_S | 21.28 GB | 0.023484 | 0.3466 | 1.6264 | 1.0028 | 96.31 % | 7.025 | 4.795 | 97.3 |

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

| ≈ 18 GB | | | | | | | | | |

| (New) Sangreal 19G-ICE | ~~18.82 GB~~<br>18.76 GB | ~~0.036887~~<br>0.034761 | ~~0.5614~~<br>0.5241 | ~~2.4828~~<br>2.1468 | ~~1.0159~~<br>1.0145 | ~~95.33 %~~<br>95.35 % | ~~7.192~~<br>6.871 | ~~4.241~~<br>4.228 | ~~96.1~~<br>96.3 |

| CyberTiel 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 | 4.083 | 95.5 |

| ≈ 13.5–15 GB | | | | | | | | | |

| (New) Sangreal 15G-ICE | ~~14.83 GB~~<br>14.81 GB | ~~0.088158~~<br>0.083796 | ~~1.4526~~<br>1.3755 | ~~4.7767~~<br>4.6790 | ~~1.0467~~<br>1.0427 | ~~92.91 %~~<br>92.93 % | ~~6.853~~<br>6.536 | ~~3.341~~<br>3.336 | ~~92.2~~<br>92.5 |

| UD-IQ3_XXS | 13.60 GB | 0.105845 | 1.7903 | 5.2413 | 1.0594 | 91.96 % | 5.489 | 3.064 | 90.9 |

Table 2 — Code at long context (code.test.raw, n_ctx 16384 × 8)

Scored on positions 8,193–16,384 of each chunk (65,536 tokens, as in the 2K tables). Compare rows within

this table only. (New) rows: shipped file ~~then~~, ICE 1.6 file now.

PPL(base) 2.121504 · BF16 = 100

| 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 | 0.037584 | 0.5148 | 7.1463 | 1.0016 | 96.86 % | 7.833 | 6.080 | 96.5 |

| (New) Sangreal 25G-ICE | ~~24.85 GB~~<br>24.82 GB | ~~0.041354~~<br>0.039198 | ~~0.5975~~<br>0.5428 | ~~7.2349~~<br>6.8581 | ~~0.9993~~<br>0.9992 | ~~96.60 %~~<br>96.68 % | ~~7.686~~<br>7.385 | ~~5.599~~<br>5.593 | ~~96.2~~<br>96.4 |

| ≈ 22.5–23 GB | | | | | | | | | |

| (New) Sangreal 23G-ICE | ~~22.84 GB~~<br>22.81 GB | ~~0.045097~~<br>0.043053 | ~~0.7004~~<br>0.6588 | ~~7.2783~~<br>7.0492 | ~~0.9996~~<br>0.9995 | ~~96.39 %~~<br>96.41 % | ~~7.523~~<br>7.215 | ~~5.145~~<br>5.139 | ~~95.9~~<br>96.0 |

| UD-Q4_K_XL | 22.75 GB | 0.045938 | 0.7000 | 7.5137 | 0.9971 | 96.29 % | 7.474 | 5.126 | 95.8 |

| UD-Q4_K_M | 22.52 GB | 0.046128 | 0.6952 | 7.2629 | 0.9989 | 96.21 % | 7.130 | 5.075 | 95.8 |

| ≈ 21 GB | | | | | | | | | |

| UD-Q4_K_S | 21.28 GB | 0.051208 | 0.8028 | 8.1942 | 0.9942 | 95.97 % | 7.025 | 4.795 | 95.4 |

| Sangreal 21G-ICE | 20.85 GB | 0.051563 | 0.8315 | 7.3234 | 0.9949 | 96.01 % | 7.357 | 4.698 | 95.4 |

| ≈ 18 GB | | | | | | | | | |

| (New) Sangreal 19G-ICE | ~~18.82 GB~~<br>18.76 GB | ~~0.063182~~<br>0.058960 | ~~1.0950~~<br>0.9983 | ~~8.1007~~<br>6.9570 | ~~1.0189~~<br>1.0160 | ~~95.24 %~~<br>95.28 % | ~~7.192~~<br>6.871 | ~~4.241~~<br>4.228 | ~~94.4~~<br>94.7 |

| UD-IQ4_XS | 18.12 GB | 0.077803 | 1.4135 | 8.5959 | 1.0284 | 94.66 % | 6.757 | 4.083 | 93.3 |

| UD-Q3_K_XL | 17.23 GB | 0.084033 | 1.5944 | 8.1866 | 1.0209 | 94.21 % | — | 3.882 | 92.8 |

| ≈ 13.5–15 GB | | | | | | | | | |

| (New) Sangreal 15G-ICE | ~~14.83 GB~~<br>14.81 GB | ~~0.112369~~<br>0.107390 | ~~2.1470~~<br>2.1159 | ~~9.0985~~<br>8.3963 | ~~1.0393~~<br>1.0353 | ~~93.09 %~~<br>93.31 % | ~~6.853~~<br>6.536 | ~~3.341~~<br>3.336 | ~~90.9~~<br>91.2 |

| UD-IQ3_XXS | 13.60 GB | 0.137273 | 2.7651 | 9.3111 | 1.0729 | 92.11 % | 5.489 | 3.064 | 89.2 |

Table 3 — English text (WikiText-2)

PPL(base) 7.574505 · BF16 = 100

| 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 | 0.024330 | 0.2425 | 1.0036 | 0.9929 | 93.88 % | 7.833 | 6.080 | 96.5 |

| (New) Sangreal 25G-ICE | ~~24.85 GB~~<br>24.82 GB | ~~0.029038~~<br>0.027264 | ~~0.2823~~<br>0.2677 | ~~1.2220~~<br>1.1085 | ~~0.9857~~<br>0.9847 | ~~93.27 %~~<br>93.21 % | ~~7.686~~<br>7.385 | ~~5.599~~<br>5.593 | ~~96.0~~<br>96.1 |

| CyberTiel 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 | | | | | | | | | |

| (New) Sangreal 23G-ICE | ~~22.84 GB~~<br>22.81 GB | ~~0.033585~~<br>0.031606 | ~~0.3299~~<br>0.3093 | ~~1.3758~~<br>1.1956 | ~~0.9870~~<br>0.9907 | ~~92.69 %~~<br>92.78 % | ~~7.523~~<br>7.215 | ~~5.145~~<br>5.139 | ~~95.5~~<br>95.7 |

| CyberTiel 23G-ICE | 22.84 GB | 0.032992 | 0.3290 | 1.2449 | 0.9899 | 92.75 % | 7.523 | 5.145 | 95.6 |

| UD-Q4_K_M | 22.52 GB | 0.035692 | 0.3576 | 1.3199 | 0.9766 | 92.42 % | 7.130 | 5.075 | 95.3 |

| UD-Q4_K_XL | 22.75 GB | 0.035837 | 0.3753 | 1.3951 | 0.9761 | 92.56 % | 7.474 | 5.126 | 95.3 |

| ≈ 21 GB | | | | | | | | | |

| CyberTiel 21G-ICE | 20.85 GB | 0.039658 | 0.4136 | 1.4041 | 0.9848 | 92.02 % | 7.357 | 4.698 | 94.9 |

| Sangreal 21G-ICE | 20.85 GB | 0.039968 | 0.4156 | 1.5834 | 0.9819 | 92.03 % | 7.357 | 4.698 | 94.9 |

| UD-Q4_K_S | 21.28 GB | 0.040169 | 0.4107 | 1.5040 | 0.9808 | 91.97 % | 7.025 | 4.795 | 94.9 |

| ≈ 18 GB | | | | | | | | | |

| (New) Sangreal 19G-ICE | ~~18.82 GB~~<br>18.76 GB | ~~0.060435~~<br>0.059599 | ~~0.6098~~<br>0.5982 | ~~2.1340~~<br>2.0323 | ~~0.9948~~<br>0.9907 | ~~90.15 %~~<br>90.17 % | ~~7.192~~<br>6.871 | ~~4.241~~<br>4.228 | 93.1 |

| CyberTiel 19G-ICE | 18.82 GB | 0.060573 | 0.6220 | 2.2733 | 0.9951 | 90.15 % | 7.192 | 4.241 | 93.0 |

| UD-IQ4_XS | 18.12 GB | 0.070787 | 0.7346 | 2.6002 | 1.0441 | 89.51 % | 6.757 | 4.083 | 92.2 |

| ≈ 13.5–15 GB | | | | | | | | | |

| (New) Sangreal 15G-ICE | ~~14.83 GB~~<br>14.81 GB | ~~0.126386~~<br>0.120439 | ~~1.3329~~<br>1.2781 | ~~3.9267~~<br>3.7501 | ~~1.0027~~<br>0.9987 | ~~85.89 %~~<br>86.18 % | ~~6.853~~<br>6.536 | ~~3.341~~<br>3.336 | ~~87.9~~<br>88.3 |

| UD-IQ3_XXS | 13.60 GB | 0.151623 | 1.6338 | 4.3641 | 1.0920 | 84.85 % | 5.489 | 3.064 | 86.2 |

*overall = 0.70/(1 + meanKLD) + 0.30 sameTop1, ×100. BF16 = 100.** Same

composite the CyberTiel card uses, so the two are directly comparable: 70% on how

close the whole output distribution stays, 30% on agreement about the argmax.

> Read the tail columns as shape, not order. 99.9% KLD is roughly the 33rd-worst

> token of 32,768 — an extreme order statistic with large sampling variance, so it

> inverts between adjacent files without meaning anything. Rank on mean KLD.

>

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

> is worth ~3% of a bit in attention, the shared expert or the output head. `active

> bpw weights by that; file bpw` is just size ÷ parameters. It is why a 22.81 GB

> file computes at ~7.2 bpw.

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

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

>

> Every row is measured against this lineage's own BF16 master — the huihui

> abliterated checkpoint — in one session, one binary, one reference. Quantizations

> cut from a different trunk are deliberately absent: scoring them here would

> measure the distance between trunks and call it quantization damage.

What Sangreal is

The calibration corpus these tiers were quantized against. **Built from primary

sources, not assembled from eaddario's set, bartowski's calibration_datav5, or any

other ready-made calibration file.** 77 buckets, 9,533 documents, 47,443,549 tokens,

sha256 85a6b823a6762bf6….

The full render is 92,663 chunks. 15G, 19G, 23G and 25G use a 36,025-chunk pass over it —

63x the stock Ornith imatrix (573) and 45x bartowski's calibration_datav5 (~800). 21G keeps the

earlier 17,225-chunk pass.

| block | buckets | share | documents | synthetic |

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

| code + cyber + spec | 36 | 45.10% | 5,858 | 0.1% |

| reasoning | 5 | 13.16% | 553 | 100.0% |

| agentic | 7 | 11.95% | 1,018 | 65.2% |

| science + domain | 11 | 11.24% | 1,183 | 0.3% |

| language | 11 | 9.66% | 596 | 0.0% |

| mathematics | 7 | 8.89% | 325 | 0.0% |

| total | 77 | 100% | 9,533 | 21.7% |

Split: technical 70.2 · science+domain 20.1 · language 9.7.

Synthetic is 21.7% by bytes and sits in exactly three buckets — teacher:k2-horizon,

teacher:deepseek-v4-pro, teacher:hermes. Everything outside reasoning and

agentic is real text.

Language — 11 languages, 5 non-Latin scripts

| bucket | share | docs | | bucket | share | docs |

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

| arabic | 0.92% | 48 | | russian | 0.87% | 58 |

| chinese | 0.89% | 31 | | spanish | 0.87% | 72 |

| hindi | 0.88% | 44 | | portuguese | 0.87% | 70 |

| japanese | 0.88% | 33 | | turkish | 0.87% | 56 |

| french | 0.88% | 60 | | german | 0.86% | 73 |

| english | 0.87% | 51 | | | | |

The spread across the whole block is 0.06 points, and that is deliberate:

language's job here is activation, not ranking. Each bucket sits just above the

saturation floor at 44.9% consumption, so these buckets select for the first time —

the legacy v1 spec had consumed 96% of turkish and 92% of english.

Mathematics — absent from the pool until this cycle

| bucket | share | docs |

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

| math (general) | 3.16% | 170 |

| analysis | 1.24% | 28 |

| algebra | 1.16% | 46 |

| geometry | 1.12% | 30 |

| proof | 1.08% | 28 |

| topology | 1.05% | 20 |

| number-theory | 0.06% | 3 |

Sourced from the LaTeX algebraic geometry, mathlib4, UniMath,

math-comp, open-web-math, AutoMathText and proof-pile.

The imatrix is published here — imatrix/sangreal-36025.imatrix.gguf, the exact file the ICE 1.6

tiers were quantized against. Pass it to llama-quantize --imatrix to rebuild them from the BF16 master,

or to cut your own.

What's inside

MTPv2 head, native. huihui's abliterated checkpoint already carries

Ornith-1.5's trained multi-token-prediction head; its mtp.* norms are bit-identical

to ornith-ai's on all seven tensors. Every tier ships blk.40 with nextn.* intact,

so llama.cpp can use it as a draft model for speculative decoding. ICE 1.6 tiers carry it at Q2_K experts /

Q4_K matrices; acceptance is unchanged (15G: 94.1% at --spec-draft-p-min 0.75).

Chat template: peculiar-ragdoll's Qwen-Sharp v22.5.0, embedded in the ICE 1.5 tiers. It is a genuinely

good template. The ICE 1.6 tiers embed v22.5.1 (even more sharpened): an agent-mode directive when tools

are present (off via chat_template_kwargs: {"decisive": false}) and fixed error coaching; plain chat

renders as v22.5.0.

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

error travels, not by how large the activations are. Routers stay F32, SSM decay

gates stay F32, the KV-cached projections stay F16, the always-on dense path stays

Q8_0 — together 0.14% of the model, kept exact for ~152 MB — and the entire budget is

spent on the routed expert bank, which is 93% of the parameters but only 8-of-256

active per token. ICE 1.6 moves the output head to Q6_K: its error stays

flat from 2K to 16K instead of compounding, which is why active bpw falls while KLD improves. Method and the cases where it does not win:

gbuzhf/ICE-quantization.

Files

• measurements/ — raw llama-perplexity output for every 2K row (2026-09-19), and

measurements/2026-09-25/ for the ICE 1.6 session: all three columns, the re-measured shipped and UD files,

the two control builds, and the MTP acceptance run.

• recipes/ — the exact llama-quantize tensor map for each tier (443 rules): cfg_<T>-ICE16.txt for

15G, 19G, 23G and 25G; cfg_21G-ICE_ICEbase.txt for 21G.

• imatrix/ — sangreal-36025.imatrix.gguf (ICE 1.6 tiers: 15G, 19G, 23G and 25G; 36,025 chunks × 512).

1020 tensors. With this plus recipes/, every ICE 1.6 tier here is reproducible byte-for-byte from the

published BF16 master.

Credits

| | |

|---|---|

| base model | ornith-ai/Ornith-1.5-35B-A3B |

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

| chat template | peculiar-ragdoll/Qwen-Sharp-Chat-Templates |

| Official CyberTiel UD ladder | peculiar-ragdoll/Cyber-Tiel-Coder-35B-A3B-GGUF-MTP |

| our previous ICE ladder | gbuzhf/…-CyberTiel-Calibrated-MTPv2-ICE-GGUF |

| ICE method | gbuzhf/ICE-quantization |

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.