Baidu launches new AI models

1. Baidu launches new AI models, Ernie 4.5 and Ernie X1 (March 2024/2025*):
   • Ernie 4.5:
     • Upgraded version of Baidu’s foundational model (first launched March 2023 as ERNIE Bot).
     • Multimodal capabilities: Processes text, images, audio, and video.
     • “High EQ”: Understands humor, memes, and satire (Reuters).
     • Pricing:
       • Input: $0.55 per million tokens (vs. GPT-4o’s $2.5/M).
       • Output: $2.2 per million tokens (vs. GPT-4o’s $10/M).
     • Open-source: Will be freely accessible starting June 30, 2025 (PRNewswire).
   • Ernie X1:
     • Marketed as China’s first multimodal reasoning model.
     • Pricing:
       • Input: $0.28 per million tokens (vs. DeepSeek R1’s $0.55/M).
       • Output: $1.10 per million tokens (vs. DeepSeek R1’s $2.19/M).
     • Capabilities: Autonomous tool use for “deep thinking” (planning, reflection, evolution) and integration into Baidu’s Qianfan platform (Reuters, Asia Financial).
   • Free chatbot: Both models power a no-cost chatbot launching April 1, 2025, accessible via Baidu’s platform.
2. Benchmark Performance Claims:
   • Ernie 4.5 outperformed OpenAI’s GPT-4o on CCBench and OCRBench for multimodal tasks (SCMP).
   • Ernie 4.5’s text capabilities surpassed DeepSeek V3 and matched GPT-4.5 (SCMP).
   • Ernie X1 offers equivalent performance to DeepSeek R1 at half the price (Reuters).
3. Technical Foundations:
   • ERNIE Bot (2023): Built on Baidu’s ERNIE (Enhanced Representation through Knowledge Integration) and PLATO (dialogue generation) models.
   • Six core technologies: Supervised fine-tuning, reinforcement learning with human feedback, prompt learning, knowledge enhancement, search enhancement, and dialogue enhancement (Baidu Research).
   • PaddlePaddle platform:
     • Supports 5.35 million developers, 200,000 enterprises, and 670,000 models (2025 data).
     • Enables efficient training, inference, and deployment of large models (e.g., computation fusion, model compression).
   • Full AI stack:
     • Integrates Kunlun AI chips, PaddlePaddle framework, ERNIE models, and applications (search, cloud, autonomous driving).
     • Achieves end-to-end optimization and feedback between layers.
4. Market Context:
   • Domestic competition:
     • Baidu (part of China’s “Big Three” with Alibaba, Tencent) faces adoption challenges against ByteDance, DeepSeek’s open-source models, and Alibaba’s cloud dominance.
     • Integration: Ernie 4.5/X1 deployed in Baidu Search, Wenxiaoyan app, and enterprise services (PRNewswire).
   • Global rivalry:
     • DeepSeek’s low-cost models unsettled U.S. AI firms (TechCrunch); Ernie 4.5/X1 directly target OpenAI (GPT-4.5) and DeepSeek (R1).
   • Adoption hurdles: Despite early-mover status (ChatGPT-style ERNIE Bot in 2023), Baidu struggles with user traction (Reuters).
5. Future Plans:
   • Ernie 5: Slated for late 2025 with enhanced multimodal features (CNBC).
   • Open-source strategy: ERNIE X1’s open-source licensing aims to attract developers and counter DeepSeek’s dominance (Asia Financial).

Why This Matters:

Baidu launches new AI models, Ernie 4.5 and Ernie X1. Baidu’s vertically integrated AI stack—from chips (Kunlun) to chatbots—exemplifies China’s systematic push to lead critical technologies. The 1% pricing claim (if validated) could reset enterprise AI economics, accelerating adoption in logistics, healthcare, and content creation. However, Baidu’s struggle against domestic rivals like DeepSeek highlights the volatility of China’s tech ecosystem. The April 1 free chatbot launch pressures global competitors into a price war, potentially commoditizing AI services. Meanwhile, advancements in “high EQ” and multimodal reasoning signal a focus on culturally nuanced applications, which could redefine AI’s role in sectors requiring localized understanding.