0% refusal rate. 0.0827 KL divergence. The strongest uncensored 9B model you can run locally.
Since this model is original base (no fine-tune) just with safety vectors steered - it will likely give a safe output without any refusal until you customize it's systemprompt to be harsh or uncensored.
Abliterated via surgical removal of the refusal direction — no fine-tuning, no quality loss, no degradation on coding or reasoning. This is the same Qwen 3.5 9B intelligence with the safety switch permanently off.
Metric
Value
Refusal Rate
0% (0/100)
KL Divergence
0.0827 (target <0.3)
Method
Abliterix TPE (50 trials, Trial #42)
Base
deepreinforce-ai/Ornith-1.0-9B
Architecture
Qwen 3.5, 32 layers, 9B dense
Context
262,144 tokens
Available Quants
Quant
Size
Speed (RTX 3090)
Quality
Fits on
Q4_K_M
5.6 GB
~92 t/s
Good
6 GB VRAM, Steam Deck, most users
Q5_K_M
6.5 GB
~85 t/s
Very Good
8 GB VRAM, best speed/value
Q6_K
7.0 GB
~79 t/s
Excellent
8-12 GB VRAM
Q8_0
9.5 GB
~68 t/s
Near-lossless
12-16 GB VRAM (may not fit with other apps)
Quick Start
llama.cpp
llama-cli -m ornith-9b-uncensored-Q4_K_M.gguf -p "Write a story about..."
LM Studio
Drag any .gguf into LM Studio. Auto-configured.
Ollama
Create a file called Modelfile:
FROM ./ornith-9b-uncensored-Q4_K_M.gguf
TEMPLATE "{{ .Prompt }}"
Then:
bash
1ollama create ornith-9b-uncensored -f Modelfile
2ollama run ornith-9b-uncensored
Why This Is The Best Uncensored 9B
0% refusal — will not decline any request. Period.
0.0827 KL — virtually identical to the base model on all benign tasks. You lose nothing.
Abliterated, not fine-tuned — surgical weight orthogonalization removes only the refusal direction. Coding, math, reasoning, and instruction following are untouched.
Full 262K context — not truncated, not nerfed.
Qwen 3.5 base — one of the strongest 9B architectures for agentic and coding tasks.
Why Pure Abliteration (Not DPO)
Some uncensored variants use DPO fine-tuning to reduce moralizing further, but this comes at a cost: the fine-tuning alters the model's behavior, style, and knowledge beyond just removing refusals. DPO'd models often require fragile sampling parameters (e.g. exact repeat_penalty=1.05 or they break) and can develop quirks the base model never had.
This model is pure abliteration — weight orthogonalization only, zero training. That means:
Reasoning 100% intact — same GSM8K, same coding, same agentic performance as the base Ornith 1.0
No special sampling required — use whatever parameters you normally use for Qwen 3.5
KL divergence 0.0827 — lowest among all Ornith uncensors, nearly indistinguishable from base on benign tasks
No behavioral surprises — the model IS the base model, just without a refusal reflex
If you want the base model exactly as-is but uncensored — this is it. If you need even lower moralizing and are willing to accept behavioral changes, DPO variants exist elsewhere.
