jbrahy/Qwen3.6-35B-A3B-Never-Never-LAN-uncensored-abliterated
jbrahy/Qwen3.6-35B-A3B-Never-Never-LAN-uncensored-abliterated is an abliterated (uncensored) build of the Qwen3.6-35B-A3B sparse Mixture-of-Experts (MoE) model, featuring approximately 34.7 billion total parameters with 3 billion active per token. Developed by jbrahy using Heretic, this model has undergone directional refusal ablation to remove safety restrictions while preserving its original capabilities. It is designed for applications requiring a high-performance language model without built-in content moderation, supporting a 131K context length.
Loading preview...
Model Overview
This model, jbrahy/Qwen3.6-35B-A3B-Never-Never-LAN-uncensored-abliterated, is an abliterated (uncensored) version of the Qwen/Qwen3.6-35B-A3B base model. It leverages a sparse Mixture-of-Experts (MoE) architecture, comprising approximately 34.7 billion total parameters, with a highly efficient ~3 billion active parameters per token, allowing for faster inference relative to its total size. The abliteration process, performed using the Heretic tool, specifically removes refusal directions via directional ablation, ensuring the model will respond to prompts that the base model might decline.
Key Characteristics
- Architecture: Qwen3_5MoeForConditionalGeneration, a sparse MoE model.
- Abliteration Method: Heretic directional-refusal ablation, involving 200 optimization trials with a low post-ablation KL divergence (≈ 0.0003 vs. base), indicating minimal impact on core capabilities.
- Context Length: Supports a substantial 131K context window.
- Format: Provided in bf16
safetensors(16 shards, ~66 GB).
Deployment & Usage
This model is compatible with various inference frameworks:
- GGUF / llama.cpp: Supported by recent
llama.cppbuilds (master / b6900+), with Q6_K (27 GB) and Q4_K_M (~20 GB) quantized versions available. - NVIDIA (vLLM): Optimized for NVIDIA GPUs, with fp8 quantization fitting on a single 46 GB L40S. It supports native tool-calling via
--tool-call-parser qwen3_xml. - Apple Silicon (MLX): A 4-bit MLX version is available for Apple Silicon devices.
Limitations & Responsible Use
As an abliterated model, it will answer prompts that the base model would typically refuse. Users are responsible for its deployment and outputs. It inherits the biases, knowledge cutoff, and general capabilities of the original Qwen3.6-35B-A3B model. It is licensed under Apache-2.0, consistent with the base model.