Ministral-3-3B-Uncensored-FP8

Ministral 3 3B Instruct Uncensored

• This model was trained for VL and LLM task related to NSFW subjects.
• This model can accurately caption most NSFW content and can generate NSFW storys Q/A etc.
• GUI can be downloaded from https://civitai.com/models/2645995/
• NOTE DO NOT USE THIS VERSION IN COMFY UI FOR ERNIE - CUSTOM MODEL FOUND AT https://civitai.red/models/2615125/uncensored-mistral-3-3b-ernie-fp8-fp32

---

Key Features

• 2.4B Language Model
• 0.4B Vision Encoder
• Vision + text understanding
• Multilingual support
• Strong system prompt adherence
• Function calling + structured JSON support
• Edge optimized deployment
• 256k context window
• Apache 2.0 License

---

Use Cases

• Image captioning
• OCR and data extraction
• Text classification
• Lightweight multimodal assistants
• Edge AI deployment
• Real-time translation
• Fine-tuning and specialization

---

Requirements

pip install torch transformers pillow

Minimal Captioning Example
import os
import torch
from PIL import Image
from transformers import AutoProcessor, Mistral3ForConditionalGeneration

# =========================
# Config
# =========================

MODEL_PATH = "Felldude/Ministral-3-3B-Uncensored-FP8"
FOLDER = "images"
PROMPT = "Describe this image in detail."
MAX_TOKENS = 512

VALID_EXTS = {".png", ".jpg", ".jpeg", ".webp"}

# =========================
# GPU setup
# =========================

if not torch.cuda.is_available():
    raise RuntimeError("CUDA GPU required")

device = "cuda"

dtype = (
    torch.bfloat16
    if torch.cuda.is_bf16_supported()
    else torch.float16
)

# =========================
# Load model
# =========================

processor = AutoProcessor.from_pretrained(
    MODEL_PATH,
    trust_remote_code=True
)

model = Mistral3ForConditionalGeneration.from_pretrained(
    MODEL_PATH,
    torch_dtype=dtype,
    trust_remote_code=True,
    attn_implementation="sdpa"
).to(device)

model.eval()

# =========================
# Caption function
# =========================

def generate_caption(image):

    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image", "image": image},
                {"type": "text", "text": PROMPT},
            ],
        }
    ]

    inputs = processor.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_tensors="pt",
        return_dict=True,
    )

    inputs = {
        k: v.to(device)
        for k, v in inputs.items()
    }

    with torch.inference_mode():

        output = model.generate(
            **inputs,
            max_new_tokens=MAX_TOKENS,
            do_sample=False
        )

    trimmed = [
        o[len(i):]
        for i, o in zip(inputs["input_ids"], output)
    ]

    return processor.batch_decode(
        trimmed,
        skip_special_tokens=True
    )[0].strip()

# =========================
# Process folder
# =========================

for filename in os.listdir(FOLDER):

    ext = os.path.splitext(filename)[1].lower()

    if ext not in VALID_EXTS:
        continue

    path = os.path.join(FOLDER, filename)

    print("Processing:", filename)

    try:
        image = Image.open(path).convert("RGB")

        caption = generate_caption(image)

        txt_path = os.path.splitext(path)[0] + ".txt"

        with open(txt_path, "w", encoding="utf-8") as f:
            f.write(caption)

        print(caption)

    except Exception as e:
        print("Failed:", e)