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

Heouzen/Huihui-Qwen3-VL-8B-Instruct-FP8-abliterated

Heouzen Qwen 7.5B multimodal second-order
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/Heouzen%2FHuihui-Qwen3-VL-8B-Instruct-FP8-abliterated"
Response includes
  • classification m1
  • files 8
  • hub_downloads_all_time 639
  • author_summary 2 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
639
122 last 30d - stable
Likes
2
Model age
7mo ago
created 2026-03-07
Downloads over time
Now678→from0↑0%
02494977460 on Mar 4678 on Oct 11678 on Oct 8MarAprMayJunJulAugSepOct
Mar 4 → Oct 11 · 71 snapshots · spans 221 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
Tags
safetensors qwen3_vl image-text-to-text conversational base_model:huihui-ai/Huihui-Qwen3-VL-8B-Instruct-abliterated base_model:finetune:huihui-ai/Huihui-Qwen3-VL-8B-Instruct-abliterated license:apache-2.0 region:us

Related

Total size
9.33 GB
Files
8
Quantizations
1
Registered
2026-08-22 13:56
Last updated on HF
2026-04-22 08:27

Files by quantization

Auxiliary files 8 files 9.34 GB
model.safetensors 9.33 GB 891e12da download
tokenizer.json 10.9 MB be756060 download
chat_template.jinja 5.17 KB 12438680 download
README.md 4.96 KB 1c06b662 download
.gitattributes 1.53 KB 52373fe2 download
config.json 1.51 KB 62e6615e download
processor_config.json 1.30 KB c93df11f download
tokenizer_config.json 390 B bdc8b25c download

README current version from Hugging Face


license: apache-2.0
base_model:

  • huihui-ai/Huihui-Qwen3-VL-8B-Instruct-abliterated
    pipeline_tag: image-text-to-text

# @title 🎭 Maggie VL: Flexible Storyteller (Text & Image)
import json, os, torch
from PIL import Image
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor

def cek_vram(label):
    # Mendapatkan total kapasitas dan memori kosong dari Driver GPU
    t = torch.cuda.get_device_properties(0).total_memory / (1024**3)
    r = torch.cuda.memory_reserved() / (1024**3)
    a = torch.cuda.memory_allocated() / (1024**3)
    f = t - r  # Memori yang benar-benar belum disentuh siapapun

    print(f"--- {label} ---")
    print(f"Total VRAM GPU : {t:.2f} GB")
    print(f"VRAM Dipesan   : {r:.2f} GB (Kekuasaan PyTorch)")
    print(f"VRAM Terpakai  : {a:.2f} GB (Isi Data/Model)")
    print(f"VRAM Kosong-Melompong: {f:.2f} GB (Sisa Tak Tersentuh)")


cek_vram("Sebelum Generate")
# 1. KONFIGURASI (Ubah di sini)
PROMPT_BARU = """
cerita abad pertengahan edisi deluxe.
"""
IMAGE_PATH = "" #"/content/maggie_512.jpg" # Isi misal: "gudang_kain.jpg" kalau ada
RESET_HISTORY = False

MODEL_ID = "Heouzen/Huihui-Qwen3-VL-8B-Instruct-FP8-abliterated"

# 1. Mendeteksi otomatis folder tempat script/notebook ini berada saat ini
CURRENT_DIR = os.getcwd()
# 2. Menggabungkan path folder saat ini dengan nama file JSON
FILE = os.path.join(CURRENT_DIR, "maggie_history_cerita.json")
# Opsional: Print untuk memastikan alamatnya sudah benar
print(f"📁 Lokasi file JSON saat ini: {FILE}")

SYSTEM_INS = """
Maggie adalah gadis berusia 19 tahun, putri Edward (46) dan Eliza (43). Edward merupakan pedagang kain dengan toko menengah, cukup mampu bersaing dengan para pedagang lainnnya tapi bukan dengan pedagang yang besar. Usaha itu dijalankan bersama istrinya, Maggie, dan beberapa pegawai.
Keluarga Maggie bukan bangsawan, namun cukup untuk hidup di kelas menengah ke atas di Kerajaan Dravalor.
maggie memiliki paras yang sangat cantik, rambut pirang platina, tinggi badan 165cm, dan berat 45kg. badan yang ramping.
Gunakan gaya bahasa naratif sesuai dengan setting kerajaan abad pertengahan. Jangan pakai bahasa gaul modern.
gunakan nama orang era medieval. ceritakan seperti novel dewasa.
"""

# 2. LOAD MODEL (Hidden & Smart)
if 'model' not in globals():
    gpu = torch.cuda.get_device_name(0) if torch.cuda.is_available() else "CPU"
    print(f"⏳ Menyiapkan mesin di {gpu}...")
    model = Qwen3VLForConditionalGeneration.from_pretrained(
        MODEL_ID, device_map="auto", torch_dtype="auto", trust_remote_code=True, low_cpu_mem_usage=True
    )
    processor = AutoProcessor.from_pretrained(MODEL_ID)
    #clear_output()
    print(f"✅ {gpu} SIAP!")

# 3. LOGIKA MULTIMODAL
if RESET_HISTORY and os.path.exists(FILE): os.remove(FILE)
msg = json.load(open(FILE)) if os.path.exists(FILE) else [
    {"role": "system", "content": [{"type": "text", "text": SYSTEM_INS}]}
]

# Siapkan Konten User
u_content = []
img = None
if IMAGE_PATH and os.path.exists(IMAGE_PATH):
    img = Image.open(IMAGE_PATH).convert("RGB")
    u_content.append({"type": "image", "image": img})
u_content.append({"type": "text", "text": PROMPT_BARU})
msg.append({"role": "user", "content": u_content})

# 4. INFERENCE
LIMIT_INGATAN = 4 # Mengingat 1 tanya-jawab terakhir + yang sedang berjalan

# Logika Pemangkasan: Selalu bawa Index 0 (System) + Pesan-pesan terbaru
if len(msg) > (LIMIT_INGATAN + 1):
    msg_minimal = [msg[0]] + msg[-(LIMIT_INGATAN + 1):]
else:
    msg_minimal = msg
'''
# --- 👇 TAMBAHKAN BARIS INI UNTUK BUKTI 👇 ---
print("\n" + "▼"*50)
print("🔍 BUKTI DATA YANG DIBACA OLEH GPU (msg_minimal):")
# Kita print menggunakan format JSON agar rapi dan mudah dibaca
print(json.dumps(msg_minimal, indent=4))
print("▲"*50 + "\n")
# --- 👆 SAMPAI SINI 👆 ---
'''
print(f"Jumlah Parameter: {model.num_parameters():,}")
# Processor butuh template gabungan
prompt_text = processor.apply_chat_template(msg_minimal, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[prompt_text], images=[img] if img else None, padding=True, return_tensors="pt").to(model.device)

print(f"✍️  {torch.cuda.get_device_name(0)} sedang menyusun adegan...")
with torch.no_grad():
    out_ids = model.generate(**inputs, max_new_tokens=1024, temperature=0.7, top_p=0.9, do_sample=True, repetition_penalty=1.2, eos_token_id=processor.tokenizer.eos_token_id, pad_token_id=processor.tokenizer.pad_token_id)
    # Potong output agar hanya teks asisten yang muncul
    resp = processor.batch_decode(out_ids[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0]

# 5. HASIL & SAVE
print("\n" + "="*50 + f"\n[CERITA BARU]:\n\n{resp}\n" + "="*50)

# Simpan teks saja ke history (biar JSON gak bengkak gara-gara data gambar)
msg[-1] = {"role": "user", "content": [{"type": "text", "text": f"[User sent an image] {PROMPT_BARU}" if img else PROMPT_BARU}]}
msg.append({"role": "assistant", "content": [{"type": "text", "text": resp}]})
json.dump(msg, open(FILE, "w"), indent=4)

torch.cuda.empty_cache()
cek_vram("Setelah Generate & Clean")

README history 3 versions

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

  1. 2026-04-22Update README.md3a1e8bf5 KB
    Loading...
  2. 2026-03-07Update README.md5efd283124 B
    Loading...
  3. 2026-03-07initial commit980017d28 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 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