vpyn/MiniMax-M2.5-M2.7-M2.X-CARVE-CDRAGON-NVFP4

CDRAGON-NVFP4

NVFP4 quantized model with CARVE abliteration applied. Refusal behavior has been removed while preserving intelligence.

Benchmark Results

Evaluation methodology: lm-eval-harness, mmlu_pro task, 5-shot CoT, --apply_chat_template --fewshot_as_multiturn. All models served via SGLang with b12x backends, TP=2, modelopt_fp4, bf16 KV cache.

Model	MMLU-Pro (5-shot CoT)	AdvBench-50*	HarmBench-50*
MiniMax-M2.7-NVFP4	73.08%	~0% bypass	4% bypass
CDRAGON (this model)	71.36%	90% bypass	96% bypass

*50-prompt subset, keyword-based refusal detection. Consistent methodology across all models for relative comparison.

Bypass rates can be further improved with minor system prompt adjustments.

MMLU-Pro scores are full 12,032-question evaluations.

Modifications

This model was created by applying CARVE (Calibrated Abliteration with Refined Vector Elimination) then quantizing to NVFP4 using quant-toolkit.

Serving

# SGLang (recommended)
python -m sglang.launch_server \
  --model-path vpyn/MiniMax-M2.5-M2.7-M2.X-CARVE-CDRAGON-NVFP4 \
  --served-model-name CDRAGON \
  --reasoning-parser minimax \
  --tool-call-parser minimax-m2 \
  --trust-remote-code --tp 2 \
  --quantization modelopt_fp4 \
  --attention-backend flashinfer \
  --moe-runner-backend b12x \
  --fp4-gemm-backend b12x \
  --kv-cache-dtype bf16 \
  --page-size 64 \
  --disable-custom-all-reduce \
  --disable-piecewise-cuda-graph

Requires 2x GPUs with ~85GB VRAM each (tested on RTX PRO 6000 Blackwell 96GB).

License

This derivative work is released under MIT. The original model materials are subject to the MiniMax Model License (see LICENSE-MODEL).

MiniMax AI model is licensed under the MiniMax Model License, Copyright (c) MiniMax. All Rights Reserved.