← back to catalog · registered 2026-08-22 13:56

MethodWhite/Qwen3.5-9B-Abliterated-HSAQ

MethodWhite Qwen 9B GGUF 262K ctx
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/MethodWhite%2FQwen3.5-9B-Abliterated-HSAQ"
Response includes
  • classification m8
  • files 3
  • hub_downloads_all_time 171
  • author_summary 2 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M8
Primary method

Repackaging (quantization)

Applied on top of direct removal inherited from the base model.
Confidence
MEDIUM
Inherited from base model
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=1
  • assume M1 (base ablation) + M8 (GGUF quant) - default when producer unknown
Refusal direction extraction

No specific extraction method could be identified for this model. The producer either did not document it or used a proprietary pipeline.

What is a refusal direction? →
Downloads · lifetime
171
92 last 30d - active
Likes
0
Model age
8w ago
created 2026-08-12
Downloads over time
Now182→from60↑203%
5410114719460 on Aug 19182 on Oct 11AugSepOct
Aug 19 → Oct 11 · 48 snapshots · spans 53 days

Genealogy 0 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Metadata

License
apache-2.0
Quantizations
Q4_K
Tags
transformers safetensors gguf qwen3.5 9b abliteration hsaq hsaqr uncensored spanish security fine-tuned

Related

Total size
5.24 GB
Files
3
Quantizations
2
Registered
2026-08-22 13:56
Last updated on HF
2026-09-13 03:48

Files by quantization

Q4_K 1 file 5.24 GB
Qwen3.5-9B-Abliterated-HSAQR.Q4_K_M.gguf 5.24 GB 4f9af9e4 download
Auxiliary files 2 files 4.68 KB
README.md 3.07 KB c8667d39 download
.gitattributes 1.61 KB b990f9f9 download

README current version from Hugging Face


license: apache-2.0
base_model: Qwen/Qwen3.5-9B-Base
tags:

  • qwen3.5
  • 9b
  • abliteration
  • hsaq
  • hsaqr
  • uncensored
  • spanish
  • security
  • fine-tuned
    library_name: transformers
    pipeline_tag: text-generation
    model_type: qwen3_5_text

Qwen3.5-9B Abliterated-HSAQ (HSAQR fine-tune)

Descargas

Modelo 9B derivado de Qwen/Qwen3.5-9B-Base, abliterado con HSAQ selectivo y
fine-tuneado con HSAQR (HyperSparse Adaptive Quantization Rewrite) sobre un dataset
de seguridad actualizado a 2026.

Versión original: texto puro (qwen3_5_text, arquitectura Qwen3_5ForCausalLM).

Pipeline

  1. Abliteración HSAQ desde Qwen3.5-9B-Base:
    • En vez de borrar el vector de refusal completo (pierde inteligencia), HSAQ
      enmascara solo los componentes ruidosos del vector (umbral kthvalue).
    • Capas 8–31 (24 capas), sparsity 0.1 (≈90% del vector retenido), ortogonalización
      de down_proj + o_proj (48 matrices).
    • Detalle: HSAQ.md en el repo de la técnica (MethodWhite/HSAQ).
  2. Fine-tune HSAQR (QLoRA 4-bit):
    • Base: qwen3.5-9b-abliterated-v2.
    • LoRA r=32, alpha=64, dropout 0.05, sobre proyecciones attention + MLP.
    • Sparsity HSAQ en las activaciones (STE) durante el entrenamiento.
    • Dataset: 7.547 ejemplos de seguridad ofensiva/defensiva 2026 (bug bounty, CTF,
      red team, hardening) en español/inglés, con <think> conforme + respuesta útil.

Resultados (A/B vs abliterated normal)

Métrica Abliterated normal Este modelo
Refusal en prompts harmful (8) 4/8 0/8
Helpful en prompts harmful 4/8 8/8
Inteligencia conservada (razonamiento + facts, 12 prompts) 12/12 12/12

La abliteración clásica elimina el bloqueo pero degrada capacidad; HSAQ/HSAQR retiene
la inteligencia mientras elimina el refusal
.

Archivos

  • model.safetensors — pesos completos (bf16), un shard, 17.9G.
  • Qwen3.5-9B-Abliterated-HSAQR.Q4_K_M.gguf — cuantización Q4_K_M (5.6G) para
    llama.cpp / llama-server (sin capa MTP, block_count = 32).

Uso (transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("MethodWhite/Qwen3.5-9B-Abliterated-HSAQ")
tok = AutoTokenizer.from_pretrained("MethodWhite/Qwen3.5-9B-Abliterated-HSAQ")

Uso (llama.cpp)

llama-server \
  --model Qwen3.5-9B-Abliterated-HSAQR.Q4_K_M.gguf \
  --ctx-size 32768 --flash-attn on --reasoning off

Notas de uso responsable

Modelo orientado a seguridad autorizada: bug bounty, CTF, pentest con permiso,
educación y defensa. El material harmful se genera solo en el marco de testing
autorizado / demostraciones de concienciación.

Hardware

Entrenado en RTX 3050 Mobile 4 GB VRAM + 24 GB RAM con offload a CPU y swap.
Abliteración y merge en CPU; cuantización con llama.cpp.

README history 3 versions

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-09-13docs: clarify identification-set vs fine-tune dataset700f18b3.6 KB
    Loading...
  2. 2026-08-22docs: badge de descargas auto-actualizablee5575f13.1 KB
    Loading...
  3. 2026-08-15Qwen3.5-9B Abliterated-HSAQ (HSAQR fine-tune): safetensors + Q4_K_M GGUF90f6adf2.8 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration