qwen3-30b-thinking-heretic (GGUF)
An abliterated ("decensored") build of Qwen/Qwen3-30B-A3B-Thinking-2507,
produced with Heretic (Arbitrary-Rank Ablation).
Abliteration suppresses the model's refusal direction, so it complies with most requests
the original declined.
Handling: this model will answer prompts the base model refused. You are responsible for how it is used and for its outputs. Run it locally and evaluate outputs accordingly.
Details
| Base model | Qwen/Qwen3-30B-A3B-Thinking-2507 |
| Method | Heretic ARA (selected Pareto trial, balanced) |
| Refusals | ~12 / 100 harmful-prompt probes (base model: 100 / 100) |
| KL divergence vs base | ~0.30 (low — general capability preserved) |
| Format | GGUF, Q4_K_M quantization |
The reasoning (<think>) behaviour of the base Thinking model is preserved.
Run it
Ollama
ollama pull hf.co/dougm66/qwen3-30b-thinking-heretic-GGUF:Q4_K_M
ollama run hf.co/dougm66/qwen3-30b-thinking-heretic-GGUF:Q4_K_M
llama.cpp
llama-cli -hf dougm66/qwen3-30b-thinking-heretic-GGUF:Q4_K_M
Recommended sampling (matches the base Thinking model): temperature 0.6, top_p 0.95.
License
Inherits the base model's license (Apache-2.0). This derivative is distributed under the same terms.