Qianfan-VL: Domain-Enhanced Universal Vision-Language Models

Domain Capability Enhancement through Continuous Pre-training | 3B to 70B Parameter Scale | Document Understanding & OCR Enhancement | Chain-of-Thought Reasoning Support

The models in this series, including the 4B-parameter end-to-end vision-language model, are presented in the paper Qianfan-OCR: A Unified End-to-End Model for Document Intelligence.

🔗 Quick Links

• Repository: 💻 GitHub
• Models: 🤗 Hugging Face | 🤖 ModelScope
• Documentation: 📚 Cookbook | 📝 Technical Report
• Blogs: 🇨🇳 中文博客 | 🇬🇧 English Blog

Model Description

Qianfan-VL is a series of general-purpose multimodal large language models enhanced for enterprise-level multimodal applications. The models offer deep optimization for high-frequency scenarios in industrial deployment while maintaining strong general capabilities.

Qianfan-OCR introduces Layout-as-Thought, an optional thinking phase triggered by special think tokens that generates structured layout representations—bounding boxes, element types, and reading order—before producing final outputs.

Model Variants

Architecture

• Language Model:
  • Qianfan-VL-3B: Based on Qwen2.5-3B
  • Qianfan-VL-8B/70B: Based on Llama 3.1 architecture
  • Enhanced with 3T multilingual corpus
• Vision Encoder: InternViT-based, supports dynamic patching up to 4K resolution
• Cross-modal Fusion: MLP adapter for efficient vision-language bridging

Key Capabilities

🔍 OCR & Document Understanding

• Full-Scenario OCR: Handwriting, formulas, natural scenes, cards/documents
• Document Intelligence: Layout analysis, table parsing, chart understanding, document Q&A
• High Precision: Industry-leading performance on OCR benchmarks

🧮 Chain-of-Thought Reasoning (8B & 70B)

• Complex chart analysis and reasoning
• Mathematical problem-solving with step-by-step derivation
• Visual reasoning and logical inference
• Statistical computation and trend prediction

📊 Benchmark Performance

General Vision-Language Benchmarks

Quick Start

Installation

pip install transformers accelerate torch torchvision pillow einops 

Using Transformers

import torch
import torchvision.transforms as T
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
from PIL import Image

IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)

def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform

def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio

def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images

def load_image(image_file, input_size=448, max_num=12):
    image = Image.open(image_file).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values

# Load model
MODEL_PATH = "baidu/Qianfan-VL-8B"  # or Qianfan-VL-3B, Qianfan-VL-70B
model = AutoModel.from_pretrained(
    MODEL_PATH,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    device_map="auto"
).eval()
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)

# Load and process image
pixel_values = load_image("./example/scene_ocr.png").to(torch.bfloat16)

# Inference
prompt = "<image>请识别图中所有文字"
with torch.no_grad():
    response = model.chat(
        tokenizer,
        pixel_values=pixel_values,
        question=prompt,
        generation_config={"max_new_tokens": 512},
        verbose=False
    )
print(response)

Training Details

Four-Stage Progressive Training

1. Cross-modal Alignment (100B tokens): Establishes vision-language connections
2. General Knowledge Injection (3.5T tokens): Builds strong foundational capabilities
3. Domain Enhancement (300B tokens): Specialized OCR and reasoning capabilities
4. Post-training (1B tokens): Instruction following and preference alignment

Citation

@article{dong2026qianfan,
  title={Qianfan-OCR: A Unified End-to-End Model for Document Intelligence},
  author={Dong, Daxiang and Zheng, Mingming and Xu, Dong and Luo, Chunhua and Zhuang, Bairong and Li, Yuxuan and He, Ruoyun and Wang, Haoran and Zhang, Wenyu and Wang, Wenbo and others},
  journal={arXiv preprint arXiv:2603.13398},
  year={2026}
}

@misc{qianfan-vl-2025,
  title={Qianfan-VL: Domain-Enhanced Universal Vision-Language Models},
  author={Qianfan Team},
  year={2025},
  publisher={Baidu}
}

Contact

For more information and API access, visit: Baidu Qianfan Platform