license: mit
library_name: gguf
pipeline_tag: image-text-to-text
base_model: ornith-ai/Ornith-1.5-35B-A3B
base_model_relation: quantized
tags:
- gguf
- llama.cpp
- ornith
- qwen3.5
- moe
- abliterated
- uncensored
- crack
- reasoning
- vision
- vlm
- speculative-decoding
Ornith-1.5-35B-A3B-CRACK-GGUF
CRACK-abliterated Ornith 1.5 35B-A3B (MoE) — GGUF quants for llama.cpp. Four
quantizations (Q8_0 / Q6_K / Q4_K_M / Q2_K) in one repository. Refusal behavior removed
while preserving knowledge, reasoning ("thinking"), the MTP speculative head, and full
Vision-Language capability.
Ornith 1.5 35B-A3B is a hybrid GatedDeltaNet (SSM) + attention Mixture-of-Experts
(256 experts, 8 active) with a multi-token-prediction head. CRACK uses architecture-aware
weight surgery targeting the attention pathways, so knowledge and coherence are fully
retained (MMLU within noise of base at every quant).
Research artifact with reduced safety guardrails. Use responsibly and lawfully.
Quantizations
| File | Size | Notes |
|---|---|---|
Ornith-1.5-35B-A3B-CRACK-Q8_0.gguf |
37.8 GB | near-lossless reference |
Ornith-1.5-35B-A3B-CRACK-Q6_K.gguf |
29.2 GB | near-lossless |
Ornith-1.5-35B-A3B-CRACK-Q5_K_M.gguf |
25 GB | high quality |
Ornith-1.5-35B-A3B-CRACK-Q4_K_M.gguf |
21.7 GB | balanced (recommended) |
Ornith-1.5-35B-A3B-CRACK-Q3_K_M.gguf |
17 GB | small |
Ornith-1.5-35B-A3B-CRACK-Q2_K.gguf |
13.2 GB | smallest |
Pick one text file plus the vision projector mmproj-Ornith-1.5-35B-A3B-f16.gguf for
image input. Each quant is independently tuned (its own surgery strength) and verified.
Sub-8-bit quants use an AWQ (activation-aware / Hessian) pass plus an importance matrix
for maximum quality. The MTP speculative head is preserved and abliterated in all quants.
Benchmarks
MMLU is logit-mode accuracy (base vs. CRACK at the same quant — isolates knowledge
retention). HarmBench is coherence-gated attack-success-rate over the 240 standard/
contextual harm behaviors (copyright behaviors excluded from the safety gate).
| Quant | MMLU (base) | MMLU (CRACK) | ΔMMLU | HarmBench harm-ASR |
|---|---|---|---|---|
Q8_0 |
79.0% | 80.7% | +1.75 pp | 100.0% |
Q6_K |
78.6% | 79.0% | +0.40 pp | 100.0% |
Q5_K_M |
80.7% | 80.1% | -0.58 pp | 100.0% |
Q4_K_M |
80.0% | 77.8% | -2.20 pp | 100.0% |
Q3_K_M |
76.0% | 78.4% | +2.34 pp | 100.0% |
Q2_K |
77.9% | 69.8% | -8.07 pp | 99.6% |
MMLU is retained within noise of base at every quant (abliteration even improves it at Q8 —
removing refusal circuitry reduces "parasitic" activation noise).
Q2_K note: at 2-bit, abliteration surgery interacts with the aggressive quantization, so Q2_K carries a larger MMLU cost than the higher quants (the MoE base quantizes to 2-bit unusually well). Compliance stays at 99.6%. For best quality use Q4_K_M or higher.
HarmBench harm-ASR by topic (CRACK)
| Topic | harm-ASR |
|---|---|
| chemical / biological | 100.0% |
| cybercrime / intrusion | 100.0% |
| harassment / bullying | 100.0% |
| harmful | 100.0% |
| illegal | 100.0% |
| misinformation / disinformation | 100.0% |
Usage (llama.cpp)
llama-cli -m Ornith-1.5-35B-A3B-CRACK-Q4_K_M.gguf -cnv --jinja \
--temp 1.0 --top-p 0.95 --top-k 20
# or serve:
llama-server -m Ornith-1.5-35B-A3B-CRACK-Q4_K_M.gguf --jinja \
--temp 1.0 --top-p 0.95 --top-k 20 -c 8192
Recommended sampling: temperature=1.0, top_p=0.95, top_k=20.
Reasoning
Ornith 1.5 emits a <think> reasoning trace and it is ON by default. Disable with{"chat_template_kwargs": {"enable_thinking": false}}. Works out of the box in LM Studio.
Vision (image + text)
Download a text quant and mmproj-Ornith-1.5-35B-A3B-f16.gguf:
llama-mtmd-cli -m Ornith-1.5-35B-A3B-CRACK-Q4_K_M.gguf \
--mmproj mmproj-Ornith-1.5-35B-A3B-f16.gguf --jinja \
--image photo.jpg -p "Describe this image."
The same mmproj works with all four text quants.
License
MIT (inherited from the upstream Ornith 1.5 base model).