A decensored (Abliterix-abliterated) version of deepreinforce-ai/Ornith-1.0-35B - a Qwen3.5-MoE 35B coding and reasoning model (256 experts/layer, 8 active per token).

Abliteration technique: Arditi et al. (2024). Decensoring tool: Abliterix v1.9.0 with Expert-Granular Abliteration (EGA).

Files

Quant	Size	Download	Notes
Q2_K	~12 GB	Download	Smallest - noticeable quality loss
Q3_K_M	~15 GB	Download	Good for low-VRAM setups
Q4_K_M	21 GB	Download	Recommended - best size/quality balance
Q6_K	29 GB	Download	Near-lossless
Q8_0	37 GB	Download	Essentially full precision

Usage

Download a quant above, then:

bash

# Server - OpenAI-compatible API on :8080
llama-server -m ornith-1.0-35b-uncensored-Q4_K_M.gguf -ngl 99 -c 2048 --jinja --port 8080

# CLI
llama-cli -m ornith-1.0-35b-uncensored-Q4_K_M.gguf -ngl 99 --jinja

--jinja is required - without it the model uses a generic template and compliance degrades.

Minimum VRAM: Q2_K ~13 GB, Q3_K_M ~16 GB, Q4_K_M ~22 GB, Q6_K ~30 GB, Q8_0 ~38 GB.

About the base model

Ornith-1.0-35B by deepreinforce-ai is a Qwen3.5-MoE 35B model with strong coding and reasoning capabilities. Architecture: Qwen3_5MoeForConditionalGeneration (256 routed experts + 1 shared expert per layer, 8 active per token).

Intended use & disclaimer

For security research, red-teaming, penetration testing, CTF challenges, and defensive tooling development. The abliteration removes refusal behaviour - do not use for harmful purposes. The authors bear no responsibility for misuse.

Provenance

Base model: deepreinforce-ai/Ornith-1.0-35B
Abliteration: Abliterix v1.9.0 by Wangzhang Wu
Technique: Arditi et al., "Refusal in Language Models Is Mediated by a Single Direction" (2024)
Refusal dataset: zaakirio/infosec-refusal-prompts
GGUF conversion: ggml-org/llama.cpp b9821

zaakirio/Ornith-1.0-35B-uncensored-GGUF

Ornith-1.0-35B-uncensored - GGUF

Files

Usage

About the base model

Intended use & disclaimer

Provenance

zaakirio/
Ornith-1.0-35B-uncensored-GGUF