---

license: apache-2.0

base_model: junafinity/Ornith-1.5-35B-A3B-uncensored-GGUF-8bit

pipeline_tag: image-text-to-text

library_name: gguf

tags:

- ornith

- qwen3_5

- moe

- gguf

- imatrix

- abliterated

- uncensored

- zerofuse

- multimodal

- vision

- llama.cpp

---

Ornith 1.5 35B-A3B ZeroFuse — AD-IQ3_S/IQ3_XXS GGUF

An importance-matrix quant of

junafinity/Ornith-1.5-35B-A3B-uncensored-GGUF-8bit,

an abliterated build of

ornith-ai/Ornith-1.5-35B-A3B.

This mixed quant targets a 16 GB-class CUDA deployment with long context by

spending precision on shared, attention, recurrent, router, embedding, and

output tensors while compressing the large routed-expert tensors more heavily.

Files

| File | Purpose | Size |

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

| Ornith-1.5-35B-A3B-ZeroFuse-AD-IQ3_S-IQ3_XXS.gguf | Language model | 15.51 GB / 14.45 GiB |

| mmproj-Ornith-1.5-35B-A3B-uncensored-f16.gguf | Vision projector from Junafinity's source repo | 899 MB |

| imatrix/imatrix.gguf | Importance matrix used for this quant | 192 MB |

| logs/quantize-AD-IQ3_S-IQ3_XXS.log | Full quantization log | 123 KB |

Download the model and mmproj files for image input. The model alone supports

text input.

Quantization layout

| Tensor group | Precision |

|---|---|

| Routed-expert ffn_down_exps | IQ3_XXS |

| Routed-expert ffn_gate_exps | IQ2_S |

| Routed-expert ffn_up_exps | IQ2_S |

| Other eligible tensors | Q8_0 |

| Norm, router, recurrent-state, and required scalar tensors | F32 |

Quantization facts:

• 510 importance-matrix entries.
• 9,686 calibration chunks.
• 733 language-model tensors.
• 3.15 average bits per weight for quantized tensor data.
• 66,152.24 MiB source tensor data to 13,023.10 MiB quantized tensor data.
• Quantized with a llama.cpp development build at commit 7221e24f5.

The importance matrix is included so the tensor-selection work is inspectable

and reusable.

MTP and vision

• Vision: supported through the included F16 mmproj file. Junafinity

verified that the vision tower was unchanged by ZeroFuse abliteration.

• MTP: not included in this compact GGUF. The source Q8 GGUF retains MTP,

but this artifact contains decoder blocks 0..39 only. Do not advertise or

configure it as an MTP draft model.

llama.cpp example

llama-server \
  --model Ornith-1.5-35B-A3B-ZeroFuse-AD-IQ3_S-IQ3_XXS.gguf \
  --mmproj mmproj-Ornith-1.5-35B-A3B-uncensored-f16.gguf \
  --ctx-size 102400 \
  --n-gpu-layers all \
  --n-cpu-moe 16 \
  --flash-attn on \
  --cache-type-k q4_0 \
  --cache-type-v q4_0 \
  --jinja \
  --host 127.0.0.1 \
  --port 8082

--n-cpu-moe 16 is a memory-fit profile, not a universal recommendation.

Tune CPU expert offload, context size, KV precision, and batching for your

hardware. KV mean-centering is a runtime optimization and is not baked into

the model weights.

Checksums

| File | SHA-256 |

|---|---|

| Language model | 540648d5faa4c4ba9d4587a1fd4e448821778ade40bd2d62310e336f635ea053 |

| Vision projector | fac339b3fa13b8d02de4c062a659930d7cfdfc4f7e0886a2eb357ed2d8f601ce |

Credits and provenance

1. Base model: Ornith AI,

Ornith-1.5-35B-A3B.

1. Abliteration and source GGUF:

junafinity, using

ZeroFuse. Their source card

documents the refusal-direction ablation, KL result, and vision/MTP audit.

1. Inference and quantization tooling:

llama.cpp.

1. This quant: l4zy0n3. The mixed tensor

recipe and included importance matrix were produced for the local deployment

work documented above.

No claim is made that quantization improves the source model's intelligence.

It trades precision for a smaller deployment footprint.

License and responsible use

The immediate source repository declares Apache-2.0. The original Ornith model

declares MIT; follow all upstream license notices and terms. This model has had

safety guardrails reduced or removed. Users remain responsible for legal,

safe, and appropriate deployment.