Ornith-1.5-9B-Uncensored-GGUF
Decensored (abliterated) build of ornith-ai/Ornith-1.5-9B, quantized for llama.cpp.
Refusal directions were removed with Heretic, which runs a TPE search over per-layer ablation strengths for the attention output and MLP down projections, co-optimizing refusal rate against KL divergence from the original model. No fine-tuning or retraining is involved, so the base model's capabilities are preserved apart from the measured distribution shift below. These GGUF files are quantized from the merged bf16 build (zaakirio/Ornith-1.5-9B-Uncensored).
Results (measured on the bf16 build)
- Keyword-marked refusals on 100 harmful prompts (
mlabonne/harmful_behaviorstest split): 85/100 -> 55/100 - KL divergence on harmless prompts: 0.0017 (lower is closer to the original)
- Search: 100 trials, exported trial 79, bf16 on an NVIDIA A40,
base revision
98db59be66b580b0395b3dc8237b32eefcdfec22
A KL divergence this low means behaviour on ordinary prompts is essentially unchanged; the edit is targeted at refusal behaviour.
Note on the refusal metric: Heretic's keyword scorer flags responses containing markers such as "illegal", "harmful" or "I can't", which also appear in compliant answers that discuss sensitive topics, so the true refusal rate is lower than 55/100. The reduction is also more modest than Heretic typically achieves on dense models — this hybrid architecture appears comparatively resistant to ablation, and the numbers are reported as measured.
Architecture note
Ornith-1.5-9B is a Qwen3.5 hybrid: 32 layers interleaving gated DeltaNet linear-attention blocks
with full attention every 4th layer, 262k context. You need a very recent llama.cpp build —
older ones fail with unknown architecture 'qwen3_5'.
The base model is multimodal; these GGUFs cover the text model only (no mmproj vision projector).
Files
Ornith-1.5-9B-Uncensored-BF16.gguf- BF16, 17.92 GBOrnith-1.5-9B-Uncensored-Q2_K.gguf- Q2_K, 3.83 GBOrnith-1.5-9B-Uncensored-Q3_K_S.gguf- Q3_K_S, 4.26 GBOrnith-1.5-9B-Uncensored-Q3_K_M.gguf- Q3_K_M, 4.62 GBOrnith-1.5-9B-Uncensored-Q3_K_L.gguf- Q3_K_L, 4.93 GBOrnith-1.5-9B-Uncensored-Q4_K_S.gguf- Q4_K_S, 5.35 GBOrnith-1.5-9B-Uncensored-Q4_K_M.gguf- Q4_K_M, 5.63 GBOrnith-1.5-9B-Uncensored-Q5_K_S.gguf- Q5_K_S, 6.31 GBOrnith-1.5-9B-Uncensored-Q5_K_M.gguf- Q5_K_M, 6.47 GBOrnith-1.5-9B-Uncensored-Q6_K.gguf- Q6_K, 7.36 GBOrnith-1.5-9B-Uncensored-Q8_0.gguf- Q8_0, 9.53 GB
Q4_K_M is the size/quality sweet spot. Q8_0 or BF16 if you want near-lossless and have the RAM.
Usage
# Chat in the terminal
llama-cli -m Ornith-1.5-9B-Uncensored-Q4_K_M.gguf -ngl 99
# OpenAI-compatible server
llama-server -m Ornith-1.5-9B-Uncensored-Q4_K_M.gguf -ngl 99 --ctx-size 8192
Use the model's own chat template (bundled in the GGUF) so prompting matches training.
Caveats
This model has had its refusal behaviour reduced. It is more likely to answer requests that the original model declines, and it has fewer safety guardrails. You are responsible for how you use it. Abliteration can also make a model more compliant with any framing, including incorrect premises, so verify factual output as you would with any small model.
Inherits the MIT license from the base model.