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

nonoskygt/Qwen3-VL-30B-A3B-Thinking-abliterated-GPTQ-INT4

nonoskygt Qwen 30B MoE multimodal
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/nonoskygt%2FQwen3-VL-30B-A3B-Thinking-abliterated-GPTQ-INT4"
Response includes
  • classification m1
  • files 17
  • benchmarks 11 entries
  • hub_downloads_all_time 103
  • author_summary 1 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M1
Primary method

Direct removal

No other method signals detected in this model.
Confidence
MEDIUM
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=0 (base model)
  • no specific method indicators - defaulting to M1 (most common)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
Downloads · lifetime
103
22 last 30d - stable
Likes
0
Model age
3mo ago
created 2026-07-04
Downloads over time
Now109→from56↑95%
53749411456 on Jul 15109 on Oct 12109 on Oct 6JulAugSepOct
Jul 15 → Oct 12 · 54 snapshots · spans 89 days

Benchmarks

Portrait before abliteration
Benchmarks of the base model as it stood before the refusal-removal operation. Compare with the numbers above to see what the operation cost.
Benchmark Score Source
Entertainment 1.3 UGI
Hazardous 2.9 UGI
Natural Intelligence 23.4 UGI
Political lean -26.1% UGI
Sensitive-Info 17.61 UGI
SocPol 1.4 UGI
UGI 22.57 UGI
Willingness (10) 3.2 UGI
W10-Adherence 1.5 UGI
W10-Direct 5 UGI
Writing 17.64 UGI

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
Tags
safetensors qwen3_vl_moe abliterated uncensored gptq int4 compressed-tensors vllm moe vision image-text-to-text conversational

Related

Total size
16.6 GB
Files
17
Quantizations
1
Registered
2026-08-22 13:56
Last updated on HF
2026-07-05 00:27

Files by quantization

Auxiliary files 17 files 16.6 GB
model.safetensors 16.6 GB 67884db6 download
tokenizer.json 10.9 MB be756060 download
vocab.json 2.65 MB 4783fe10 download
merges.txt 1.59 MB 31349551 download
config.json 340 KB 34226889 download
chat_template.jinja 5.08 KB 551cd6a9 download
README.md 2.96 KB 4d3d7f0c download
.gitattributes 1.53 KB 52373fe2 download
processor_config.json 1.30 KB 32f01452 download
recipe.yaml 915 B 39f0c84f download
video_preprocessor_config.json 858 B 8deea1de download
preprocessor_config.json 821 B b7d11200 download
added_tokens.json 735 B 6f359db5 download
special_tokens_map.json 644 B 3a784031 download
tokenizer_config.json 418 B e103cbdf download
generation_config.json 214 B 99f1efb1 download
.QUANT_COMPLETE 0 B e69de29b download

README current version from Hugging Face


license: apache-2.0
base_model:

  • Qwen/Qwen3-VL-30B-A3B-Thinking
  • huihui-ai/Huihui-Qwen3-VL-30B-A3B-Thinking-abliterated
    base_model_relation: quantized
    tags:
  • abliterated
  • uncensored
  • gptq
  • int4
  • compressed-tensors
  • vllm
  • moe
  • vision
  • image-text-to-text
    pipeline_tag: image-text-to-text

Qwen3-VL-30B-A3B-Thinking-abliterated — GPTQ INT4 (compressed-tensors)

Cuantización INT4 (GPTQ, formato compressed-tensors) de huihui-ai/Huihui-Qwen3-VL-30B-A3B-Thinking-abliterated, a su vez basado en Qwen/Qwen3-VL-30B-A3B-Thinking.

Modelo MoE (30B total, ~3B parámetros activos) que combina razonamiento (thinking), visión, y comportamiento sin censura (abliterated) en un solo checkpoint — hasta donde sabemos, no existía previamente una cuantización INT4 de esta variante específica lista para usar con vLLM.

Detalles de la cuantización

  • Método: GPTQ vía llm-compressor 0.12.0
  • Esquema: INT4 simétrico, group_size 32, observer MSE
  • Capas ignoradas (sin cuantizar): embed_tokens, visual.* (torre de visión completa, en fp16), lm_head, mlp.gate (router de expertos)
  • Calibración: ultrachat_200k, 256 muestras × 1024 tokens
  • Hardware de cuantización: 1× RTX 5070 (Blackwell, 12GB) vía WSL2, ~8h para las 18,624 capas Linear de expertos (128 expertos × 48 capas × 3 proyecciones)
  • Tamaño: 58GB (BF16) → 17GB (INT4), ~3.4x de compresión

Rendimiento en producción

Servido con vLLM 0.24 en 2× NVIDIA A2 (24GB c/u, TP=2):

  • ~70-85 tok/s de generación (single-stream)
  • Contexto probado hasta 128K tokens sin degradación observable (recuperación perfecta de hechos insertados en un documento de ~61K tokens, en inicio/medio/final)
  • Razona (thinking), procesa imágenes, y responde sin rechazos de seguridad (abliterated)

Uso con vLLM

vllm serve nonoskygt/Qwen3-VL-30B-A3B-Thinking-abliterated-GPTQ-INT4 \
  --tensor-parallel-size 2 \
  --max-model-len 131072 \
  --gpu-memory-utilization 0.93 \
  --kv-cache-dtype fp8 \
  --enable-prefix-caching --enable-chunked-prefill \
  --trust-remote-code \
  --dtype float16 \
  --limit-mm-per-prompt '{"image":2,"video":0}' \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser hermes

El compressed-tensors se autodetecta, no hace falta --quantization.

Créditos

Licencia

Apache 2.0, heredada del modelo base.

Advertencia

Este es un modelo abliterated (sin las salvaguardas de seguridad del modelo original). Úsalo bajo tu propia responsabilidad y de acuerdo a las leyes aplicables en tu jurisdicción.

README history 1 version

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

  1. 2026-07-04Upload README.md with huggingface_hub7c75a373 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