mlasli/Qwen3.6-27B-abliterated
mlasli/Qwen3.6-27B-abliterated is an uncensored, agent-friendly variant of Alibaba's 27B Qwen3.6-27B dense coding model, featuring a 32768-token context length. This version has been 'abliterated' using Heretic v1.4.0 to surgically remove refusal behaviors, achieving 88% compliance with minimal impact on original capabilities (KL divergence 0.0118). It is specifically designed for security research, red-teaming, unrestricted coding agents, and creative writing where the base model's safety alignment would interfere.
Loading preview...
Qwen3.6-27B Abliterated: Uncensored Coding and Agent Model
This model is an abliterated version of Alibaba's powerful Qwen3.6-27B, a 27-billion parameter dense coding model known for matching Claude 4.5 Opus on Terminal-Bench 2.0 and outperforming larger MoE models like Qwen3.5 MoE in agentic coding tasks. The key differentiator of this variant is the removal of its refusal vectors, making it uncensored and highly suitable for applications where the base model's safety alignments might hinder legitimate operations.
Key Capabilities and Features
- Refusal-Vector-Removed: Achieves 88% compliance (only 12% refusal rate) on challenging prompts, enabling broader utility.
- Capability Preservation: Abliteration via Heretic v1.4.0 results in an extremely low KL divergence (0.0118), indicating near-identical output distribution and capability to the original Qwen3.6-27B.
- High-Performance Base: Inherits the strong coding and agentic reasoning abilities of the Qwen3.6-27B foundation.
- Context Length: Supports a substantial 32768-token context window, extensible to 1M.
- Hardware Efficiency: Operates with approximately 55 GB VRAM in BF16, 17 GB in 4-bit, and 28 GB in 8-bit precision.
Ideal Use Cases
- Security Research & Red-Teaming: Facilitates unrestricted exploration of security vulnerabilities and penetration testing scenarios.
- Unrestricted Coding Agents: Powers coding agents (e.g., OpenCode, Cline, Aider) without encountering content-based refusals.
- Creative Writing: Enables generation of diverse content without content filters.
- Experimentation: Useful for testing prompt injection, jailbreak techniques, and alignment research.