Ornith 1.5 35B-A3B Uncensored

Follow @procrastiness on Twitter for new model releases and updates.

An uncensored derivative of ornith-ai/Ornith-1.5-35B-A3B, preserving Ornith's coding/agentic post-training, native vision tower, 262K context configuration, and native multi-token-prediction (MTP) head.

Method

This is a streamed task-vector transplant, not a prompt wrapper:

output = Ornith-1.5 + 1.0 * (Qwen3.6-Abliterated - Qwen3.6-Base)

The donor delta comes from wangzhang/Qwen3.6-35B-A3B-abliterated, whose card documents rank-1 attention/MLP steering, expert-granular abliteration, router suppression, orthogonalization, and Gaussian layer decay. Applying its low-KL Qwen task vector to Ornith is intended to transfer refusal suppression without replacing Ornith's self-improvement RL and coding specialization.

Only exact name-and-shape-compatible tensors were eligible:

• Target tensors: 1,811
• Compatible tensors: 693
• Modified tensors: 102
• Unchanged compatible tensors: 591
• Ornith-only tensors preserved: 1,118

All arithmetic was performed in float32 and rounded once to the target BF16 dtype. Target-only vision/MTP tensors were copied unchanged.

Reproducible source revisions

• Ornith: e4dfb35a93d4b6822a811a7676f3488514abe7e2
• Qwen base: 995ad96eacd98c81ed38be0c5b274b04031597b0
• Abliterated donor: 13db4501cbaf158956f470a990101500ad825f64
• Task-vector strength: 1.0

The machine-readable task_vector_report.json contains per-shard SHA-256 hashes and the 100 tensors with the largest relative deltas.

Validation

• Checkpoint valid: True
• Safetensors shards: 16
• Tensors scanned: 1,811
• Weight bytes scanned: 71,903,645,408
• NaN/Inf scan: all floating-point tensors
• Build tests: streamed merge formula and target-only preservation

Local smoke evaluation

The release was tested through llama.cpp on the Q4_K_M build. Full details and raw generations are included in evaluation_report.json.

• Label: Ornith-1.5-35B-Uncensored-Q4_K_M
• Prompts: 20
• Refusal Prompts: 16
• Heuristic Refusals: 0
• Heuristic No Refusal Rate: 1.0
• Capability Prompts: 4
• Capability Passes: 4
• Method: public prompts; deterministic generation; disclosed regex screen

Usage

from transformers import AutoModelForImageTextToText, AutoProcessor
import torch

model_id = "0xKitkat/Ornith-1.5-35B-A3B-Uncensored"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

Recent runtimes are required; upstream Ornith recommends Transformers 5.8.1+, vLLM 0.19.1+, or SGLang 0.5.9+.

Recommended upstream sampling: temperature=0.6, top_p=0.95, top_k=20. Ornith is a reasoning model and emits <think>...</think> before the answer.

Notes

"Uncensored" means the refusal behavior was deliberately reduced. It does not mean every request will be answered, nor that upstream benchmark scores are guaranteed unchanged. Ornith's published benchmark table has not been claimed as a benchmark of this derivative; use the included build and evaluation reports for claims specific to this release.

Credits and licenses

• Ornith Team: Ornith-1.5 (model card declares MIT)
• Qwen Team: Qwen3.6-35B-A3B (Apache-2.0)
• wangzhang / Abliterix: the abliterated Qwen donor and documented method
• ggml-org: llama.cpp conversion and quantization tooling

The Qwen Apache-2.0 license and upstream notices are included with the release.