REDDIT · REDDIT// 4h agoMODEL RELEASE

Qwen3.6 35B Heretic drops with near-zero KLD

Qwen3.6-35B-A3B-Heretic is a decensored version of Qwen's latest 35B Mixture-of-Experts (MoE) model, engineered to maintain its high-end reasoning and coding capabilities without the restrictive safety filters. Developed using the Magnitude-Preserving Orthogonal Ablation (MPOA) method, the model achieves a remarkably low KL Divergence of 0.0015, ensuring it stays true to the original's intelligence while slashing refusal rates by nearly 90%.

// ANALYSIS

Heretic is a masterclass in surgical model abliteration, proving that safety filters can be removed without sacrificing the underlying intelligence of a high-parameter MoE architecture.

–KL Divergence of 0.0015 is remarkably low, indicating the model's weights remain statistically nearly identical to the original despite the "lobotomy" reversal.
–The 88% reduction in refusals makes it a viable choice for complex, uncensored coding and multi-turn reasoning tasks where standard Qwen models often "preach."
–With only 3B activated parameters, it delivers high-end reasoning performance while remaining runnable on consumer hardware like a single 24GB VRAM GPU.
–Preservation of the "Thinking" context is critical for developers using it as an agentic backend for repo-level refactoring.

// TAGS

qwen3.6-35b-a3b-hereticllmopen-weightsmixture-of-expertsdecensoredai-coding

DISCOVERED

4h ago

2026-04-26

PUBLISHED

4h ago

2026-04-26

RELEVANCE

9/ 10

AUTHOR

My_Unbiased_Opinion