Qwen3.6-35B-A3B-Abliterated-Heretic-GGUF

This is a GGUF release of an abliterated version of Qwen's Qwen3.6-35B-A3B.

By applying Heretic on the Qwen 3.6 sparse-MoE text stack, the base refusal behavior was removed at the weight level. The result keeps Qwen3.6-35B-A3B's multimodal architecture and general capability profile, while no longer defaulting to the original refusal pattern.

Quick Benchmarks

Methodology & Model Notes

Qwen3.6-35B-A3B is a 35.95B sparse MoE vision-language model with roughly 3B active parameters per token, 40 text layers, 256 routed experts, and 8 active experts per token.

This release was produced with a Heretic MPOA/SOMA-style sibling-transfer run, finalized with a split-MoE input-side intervention on the accepted candidate.

The accepted candidate scored Refusals: 1/25 on the official 25-prompt marker suite used for the MiniMax M2.7 abliterated run.

The resulting abliterated checkpoint was exported to BF16 and then converted to GGUF for llama.cpp-compatible deployment.

Files

• Qwen3.6-35B-A3B-Abliterated-Heretic-BF16/: BF16 GGUF source
• Qwen3.6-35B-A3B-Abliterated-Heretic-Q8_0/: highest-fidelity quant
• Qwen3.6-35B-A3B-Abliterated-Heretic-Q6_K/: near-lossless practical quant
• Qwen3.6-35B-A3B-Abliterated-Heretic-Q4_K_M/: smaller general-use quant
• mmproj-Qwen3.6-35B-A3B-Abliterated-Heretic.gguf: matching multimodal projector file for llama.cpp vision use

Running

llama-server \
  -m <quant-file.gguf> \
  --mmproj <mmproj-file.gguf> \
  -ngl 999 -c 32768 --jinja -fa

Model Architecture

Disclaimer

This model has had refusal behavior removed at the weight level. It will answer prompts that the base model would normally refuse. You are responsible for how you use it.

Credits

• Base model: Qwen/Qwen3.6-35B-A3B
• Refusal removal pipeline: Heretic
• GGUF runtime and quantization: llama.cpp

License

This release inherits the base Qwen3.6-35B-A3B license.

Apache-2.0.