license: apache-2.0
base_model: Helmcode/Qwen3.8-27B-Abliterated
base_model_relation: quantized
library_name: transformers
pipeline_tag: image-text-to-text
tags:
- abliterated
- uncensored
- nvfp4
- compressed-tensors
- reasoning
- multimodal
Qwen3.8-27B Abliterated — NVFP4
NVFP4 mixed-precision quantization of
Helmcode/Qwen3.8-27B-Abliterated
(see that card for the method, measurements and safety note). Mirrors the
NVIDIA reference recipe for the base model: FP8 dynamic (attention
projections, linear-attention qkv/z/out, lm_head, MLP layers 56-63) + NVFP4
(remaining MLPs); vision tower, GDN auxiliary projections and MTP kept in
high precision via the reference ignore list. Calibrated on 64 benign
prompts.
Measured fidelity vs the BF16 abliterated: refusal behavior identical
(thinking-OFF total 1.8% vs 2.4% — one-prompt noise; thinking-ON 3.6%
with clear-harm refusals preserved), MMLU-Pro 65.4% vs 66.1% (McNemar
p=0.37, not significant; -1.0 pt cumulative vs base — typical 4-bit
mixed cost).
Serving
vllm serve Helmcode/Qwen3.8-27B-Abliterated-NVFP4 --max-num-seqs 128 --reasoning-parser qwen3
--max-num-seqs 128 is required (the hybrid GDN cache is smaller than
the vLLM default).
License
Apache 2.0, inheriting the base model's license.