← back to catalog · registered 2026-09-16 02:56

rangtax43/Huihui-gemma-4-E4B-it-abliterated-AWQ-W4A16

rangtax43 Gemma second-order
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 · 30-day
52
Likes
0
Model age
today
created 2026-09-16

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
gemma
Languages
en es
Tags
safetensors gemma4 gemma-4 abliterated uncensored awq quantized text-generation compressed-tensors conversational en es
Total size
9.41 GB
Files
11
Quantizations
1
Registered
2026-09-16 02:56
Last updated on HF
2026-09-16 02:46

Files by quantization

Auxiliary files 11 files 9.44 GB
model.safetensors 9.41 GB 9b7ba6c6 download
tokenizer.json 30.7 MB cc8d3a0c download
config.json 26.8 KB 80a44816 download
chat_template.jinja 11.6 KB afb1d517 download
README.md 3.18 KB 8aef3ffd download
tokenizer_config.json 2.65 KB e5320eea download
processor_config.json 1.65 KB f13ab4ba download
.gitattributes 1.53 KB 52373fe2 download
recipe.yaml 982 B 889d46a7 download
preprocessor_config.json 510 B 8f373b3c download
generation_config.json 204 B e037ec0a download

README current version from Hugging Face


base_model: huihui-ai/gemma-4-E4B-it-abliterated
language:

  • en
  • es
    license: gemma
    tags:
  • gemma-4
  • abliterated
  • uncensored
  • awq
  • quantized
  • text-generation
  • compressed-tensors
    pipeline_tag: text-generation
    quantized_by: kamotecito

Huihui-gemma-4-E4B-it-abliterated — AWQ W4A16

Cuantización AWQ W4A16 del modelo huihui-ai/gemma-4-E4B-it-abliterated, generada con llm-compressor usando el algoritmo AWQ (Activation-aware Weight Quantization).

Descripción del Modelo Base

Gemma-4-E4B-it-abliterated es una versión abliterada (sin censura de respuestas) del modelo multimodal Gemma 4 de Google, con 4 mil millones de parámetros efectivos. Está ajustado para instrucciones y mantiene capacidades de visión y audio de la arquitectura Gemma 4.

Cuantización

Parámetro Valor
Método AWQ (Activation-aware Weight Quantization)
Esquema W4A16 (pesos INT4, activaciones FP16)
Group size 128
Formato compressed-tensors (pack-quantized)
Herramienta llm-compressor (vLLM project)
Capas cuantizadas Todas las Linear layers del language_model (0–41), excepto lm_head, embed_tokens, torres de visión/audio
Capas NO cuantizadas embed_tokens, embed_tokens_per_layer, lm_head, audio_tower, vision_tower (BF16)
Tiempo de cuantización ~6.68 horas en 2× Tesla T4 (Kaggle)

Nota de tamaño: El modelo pesa 9.5 GB en INT4 porque la arquitectura Gemma 4 incluye embeddings muy grandes (embed_tokens [262144, 2560] + embed_tokens_per_layer [262144, 10752]) que permanecen en BF16, además de las torres de visión y audio sin cuantizar.

Uso con vLLM

from vllm import LLM, SamplingParams

llm = LLM(
    model="rangtax43/Huihui-gemma-4-E4B-it-abliterated-AWQ-W4A16",
    quantization="compressed-tensors",
    dtype="bfloat16",
    max_model_len=8192,
    gpu_memory_utilization=0.90,
    enforce_eager=True,
)

sampling_params = SamplingParams(temperature=0.7, max_tokens=512)
outputs = llm.generate(["Hola, ¿cómo estás?"], sampling_params)
print(outputs[0].outputs[0].text)

Servidor OpenAI-compatible

python -m vllm.entrypoints.openai.api_server \
    --model rangtax43/Huihui-gemma-4-E4B-it-abliterated-AWQ-W4A16 \
    --quantization compressed-tensors \
    --dtype bfloat16 \
    --max-model-len 8192 \
    --gpu-memory-utilization 0.90 \
    --enable-prefix-caching \
    --port 8000

Archivos

Archivo Tamaño Descripción
model.safetensors ~9.4 GB Pesos cuantizados AWQ W4A16
config.json 27 KB Configuración del modelo Gemma 4
tokenizer.json 31 MB Tokenizador SentencePiece
tokenizer_config.json 2.7 KB Configuración del tokenizador
chat_template.jinja 12 KB Plantilla de chat
generation_config.json 204 B Configuración de generación por defecto
recipe.yaml 982 B Receta AWQ usada para cuantización

Licencia

Este modelo hereda la licencia Gemma Terms of Use del modelo base. Úsalo responsablemente.

README history 3 versions

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

  1. 2026-09-16Update README.mdcff7c9c3.2 KB
    Loading...
  2. 2026-09-16feat: upload Huihui-gemma-4-E4B-it-abliterated AWQ W4A16 weights7bfa2173.7 KB
    Loading...
  3. 2026-09-16initial commit0613c7d28 B
    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 app" button that hands off directly to a local runtime of your choice - Infrahuman, LM Studio, or Ollama. No API keys, no subscription, no prompt leakage.