tuanlai001/Qwopus3.6-27B-v2-heretic-MTP-NVFP4

This model is an uncensored (heretic) version of Qwopus3.6-27B-v2, quantized to NVFP4 format.

🛠️ Training & Quantization Details

·Uncensoring Tool: heretic

·Checkpoint Selection: Manually selected from 800 trials, achieving a refusal rate of 4% (4/100) and a KL Divergence of 0.01066.

·Quantization Tool: llm-compressor

·Quantization Recipe: Adapted from unsloth/Qwen3.6-27B-NVFP4, Visual blocks, MTP, and linear attention layers are retained in BF16 precision to preserve performance.

·Evaluations: Benchmarked with the same selected MMLU-PRO test set as Jackrong(original author of Qwopus3.6-27B model), results shown below:

📚MMLU-Pro Selected Subset

Model	Correct / Total	Accuracy
Qwopus3.6-27B-v2	306 / 350	87.43%
Qwopus3.6-27B-v2-heretic-NVFP4 (this model)	303 / 350	86.57%
Qwen3.6-27B	297 / 350	84.86%

Category	Qwen3.6-27B	Qwopus3.6-27B-v2	NVFP4	Delta (vs Qwopus)
Biology	96%	96%	92%	-4 pp
Business	88%	94%	92%	-2 pp
Computer Science	82%	84%	82%	-2 pp
Mathematics	90%	88%	92%	+4 pp
Physics	76%	86%	90%	+4 pp
Chemistry	74%	80%	76%	-4 pp
Health	88%	84%	82%	-2 pp

🚀 Usage

Sample command to run this model with vLLM

vllm serve /tuanlai/Qwopus3.6-27B-v2-heretic-MTP-NVFP4 \
 --trust-remote-code \
 --dtype bfloat16 \
 --enable-auto-tool-choice \
 --tool-call-parser qwen3_coder \
 --reasoning-parser qwen3 \
 --speculative-config '{"method":"mtp","num_speculative_tokens":3}' 

Any feedback is welcomed.