ERNIE 4.5: Baidu's Next-Generation Native Multimodal Foundation Model 206

What is Wenxin 4.5 Large Model

Wenxin 4.5 Large Model is Baidu's first natively multimodal large model, featuring powerful multimodal understanding, text generation, and logical reasoning capabilities. In multiple benchmark tests, it outperforms GPT-4.5 and DeepSeek-V3, achieving an average score of 79.6 points, higher than GPT-4.5's 79.14 points.

Features of Wenxin 4.5 Large Model

Native Multimodal Capability:

• Supports multimodal input and output for text, images, audio, and video. It performs excellently in multimodal understanding, capable of handling tasks such as graphical reasoning, chart analysis, and meme comprehension.

Strong Language Capability:

• Excels in text generation, logical reasoning, code understanding, and generation, achieving an average score of 79.6 points, higher than GPT-4.5's 79.14 points.

Technical Innovations:

• FlashMask Dynamic Attention Mask: Enhances long-sequence modeling capability and training efficiency, optimizing long-text processing ability.

• Multimodal Heterogeneous Expert Expansion Technology: Solves the gradient imbalance issue across different modalities, improving multimodal fusion capability.

• Spatiotemporal Dimension Representation Compression Technology: Significantly improves multimodal data training efficiency.

• Knowledge Point-Based Large-Scale Data Construction Technology: Enhances model learning efficiency and reduces hallucinations.

• Self-Feedback-Based Post-training Technology: Improves the model's ability to align with human intent.

High Cost-Performance Ratio:

• API call pricing is only 1% of GPT-4.5, with input pricing at 0.004 yuan per thousand tokens and output pricing at 0.016 yuan per thousand tokens.

Text Generation Capability:

• The generated text is more natural and fluent, with higher accuracy and creativity.

Coding Capability:

• Supports code generation and explanation, suitable for programming assistance scenarios.

Knowledge Accuracy:

• Through knowledge point-based large-scale data construction technology, model hallucinations are significantly reduced and knowledge accuracy is improved.

High EQ and Cultural Understanding:

• Can understand internet memes, cultural allusions, etc., demonstrating "high EQ," and can provide explanations in the context of cultural backgrounds.

Application Scenarios of Wenxin 4.5 Large Model

• Content Creation: Applicable to advertising, film and television, education, and other fields, supporting text creation, image generation, video understanding, and other multimodal content generation.

• Data Analysis: With powerful logical reasoning and chart analysis capabilities, it can assist enterprises in complex data analysis and decision-making.

• Intelligent Customer Service: Its multimodal interaction capability enables it to be applied in intelligent customer service, providing more natural and efficient services.

• Online Education: Assists online education platforms in answering questions and resolving doubts.

• Programming Assistance: Provides code generation and explanation services for developers, improving development efficiency.
  

How to Use Wenxin 4.5 Large Model

Individual Users

• Wenxin Large Model 4.5 has been launched on the Wenxin Yiyan official website. Users can log in, select Wenxin 4.5 in the top navigation, and experience it for free.

Developers and Enterprise Users

• Wenxin 4.5 Large Model has been launched on the Baidu Intelligent Cloud Qianfan Large Model Platform. Developers and enterprise users can access the model via API calls.

API Call Pricing for Wenxin Large Model 4.5

• Input pricing: 0.004 yuan per thousand tokens.

• Output pricing: 0.016 yuan per thousand tokens.

This pricing is approximately 1% of GPT-4.5's, offering a high cost-performance ratio.