Kiel-2-VQA

Kiel-2-VQA is a high-performance multimodal vision-language model fine-tuned from the Qwen2.5-VL architecture. It is engineered to excel in Visual Question Answering (VQA), complex document and chart analysis, dense OCR data extraction, and fine-grained visual localization. As part of the Kiel AI suite, it serves as the core intelligence engine for high-precision image understanding and multimodal chat workflows.

Model Details

• Developed by: KielTech
• Base Model: unsloth/qwen2.5-vl-7b-instruct-unsloth-bnb-4bit
• Model Type: Multimodal Causal Language Model (Qwen2_5_VLForConditionalGeneration)
• Language(s): English
• License: Apache 2.0
• Quantization: 4-bit optimized (BNB NF4) for high-throughput, memory-efficient inference
• Hugging Face Hub: kiel2/Kiel-2-VQA

Key Capabilities

• Visual Question Answering & Reasoning: Deep comprehension of photographs, diagrams, UI layouts, and illustrations to answer complex user queries.
• Document, Table & Chart Parsing: Advanced layout understanding for structured data extraction from invoices, receipts, charts, and technical documents.
• High-Accuracy OCR: Robust text recognition across multiple fonts, orientations, and scales.
• Spatial Localization: Precise identification of object coordinates and visual elements within images.

How to Get Started with the Model

You can load and run Kiel-2-VQA directly using Hugging Face transformers and bitsandbytes for memory-efficient execution:

import torch
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from PIL import Image
import requests

MODEL_ID = "kiel2/Kiel-2-VQA"

print("Loading Kiel-2-VQA processor and model...")
processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)
model.eval()

# Prepare an image and a text query
image_url = "[https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg)"
image = Image.open(requests.get(image_url, stream=True).raw)

conversation = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": "Describe the key elements and setting of this image in detail."},
        ],
    }
]

# Apply chat template
text_prompt = processor.apply_chat_template(
    conversation, tokenize=False, add_generation_prompt=True
)

# Process inputs
inputs = processor(
    text=[text_prompt],
    images=[image],
    padding=True,
    return_tensors="pt"
).to(model.device)

# Generate response
with torch.no_grad():
    output_ids = model.generate(**inputs, max_new_tokens=256)

generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(inputs.input_ids, output_ids)
]
response = processor.batch_decode(
    generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
)[0]

print("Model Response:\n", response)
Citation   
Code snippet
@misc{kiel2-Kiel-2-VQA-2026,
  author = {KielTech},
  title = {Kiel-2-VQA: High-Performance Multimodal Vision-Language Model},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face Repository},
  howpublished = {\url{[https://huggingface.co/kiel2/Kiel-2-VQA](https://huggingface.co/kiel2/Kiel-2-VQA)}}
}