Downloads · 30 days
138
16% of all-time downloads
GoldfingerCH/Qwen3.8-27B-abliterated-NVFP4
Qwen3.8-27B-abliterated-NVFP4 is a text generation model from GoldfingerCH. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as apache-2.0.
NVFP4 (compressed-tensors) checkpoint of an abliterated Qwen/Qwen3.8-27B, packaged for vLLM on NVIDIA Blackwell / DGX Spark (GB10).
Downloads · 30 days
138
16% of all-time downloads
All-time downloads
850
Public
Parameters
26.9B
18.8 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors18.8 GB · 100%
How the weights are stored.
U824.3B · 90%
From the Hugging Face model README
NVFP4 (compressed-tensors) checkpoint of an abliterated Qwen/Qwen3.8-27B, packaged for vLLM on NVIDIA Blackwell / DGX Spark (GB10).
Abliteration orthogonalizes the refusal direction out of the residual stream. It is a weight-space edit, not a fine-tune. Multi-token prediction (MTP) and the vision tower are left unchanged.
Research only: safety alignment has been substantially removed. This checkpoint is intended for interpretability, refusal-mechanism study, red-teaming, and controlled experiments — not for production or end-user deployment.
This model has had its safety alignment substantially removed via abliteration (orthogonalizing the refusal direction out of the residual stream). As a direct consequence:
By downloading or using this model you acknowledge and accept the above.
| Step | Detail |
|---|---|
| Base | Qwen/Qwen3.8-27B |
| Edit | Abliteration on transformer layers 18–51 (other layers, MTP, and vision unmodified) |
| Quantization | NVFP4 via llm-compressor oneshot; calibration ultrachat-200k (train_sft), 256 samples, seq 4096 |
| Format | compressed-tensors, ~18 GB |
| Target | vLLM on Blackwell (sm_121a) |
export CUTE_DSL_ARCH=sm_121a
vllm serve GoldfingerCH/Qwen3.8-27B-abliterated-NVFP4 \
--quantization compressed-tensors \
--host 0.0.0.0 --port 8000 \
--served-model-name Qwen3.8-27B-abliterated-NVFP4 \
--max-model-len 262144 \
--gpu-memory-utilization 0.70 \
--kv-cache-dtype fp8 \
--enable-prefix-caching \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder \
--limit-mm-per-prompt '{"image":0,"video":0}' \
--trust-remote-code \
--api-key internal
On Spark, FP8 KV cache leaves on the order of ~1.9M tokens of cache (~7× concurrency at 262k).
Apache 2.0, inherited from the base model Qwen/Qwen3.8-27B. Abliteration does not change the underlying license obligations.