Baidu Wenxin Large Model Cost Optimization Analysis Report

According to the information found,

. Here is the detailed analysis:

According to data disclosed at Baidu’s April 2024 Create AI Developer Conference [1]:

• Algorithm training efficiency increased to 5.1 times that of a year ago
• Average weekly training efficiency reached 98.8%
• Inference performance increased by 105 times
• Cost decreased by 99%

Baidu’s launched “Wanyuan” Intelligent Computing Operating System has achieved significant results in cost control [1]:

• ModelBuilder provides model routing services
  , selecting models with appropriate parameter scales to form optimal solutions,
  reducing inference costs by 30%
• The Baige AI heterogeneous computing platform is compatible with chips from different manufacturers,
  with mixed training efficiency of 97%/95% for 100 cards/1000 cards

The Wenxin Large Model uses multiple technologies to reduce operating costs [1]:

• Mixture of Experts (MoE) model
  : Does not rely on a single model; mixes large and small models according to scenario matching
• Lightweight model strategy
  : Lightweight models such as ERNIE Speed, Lite, and Tiny reduce the demand for computing resources

With the explosive growth in API calls, costs have been further diluted [3]:

• The number of Wenxin Large Model API calls increased from 50 million in 2023 to
  1.65 billion in 2024
• The number of Wenxin Yiyan users exceeded 200 million, with daily API calls exceeding 200 million

The large model industry as a whole has entered a cost reduction cycle [2]:

• At the end of 2024, Alibaba announced a price reduction of more than 80% for its large models
• Baidu Intelligent Cloud’s AI-related revenue grew by nearly 300%, and the scale effect further reduced marginal costs
• Baidu expects the cost of basic models to continue to
  decline steadily
  , lowering the threshold for use

, mainly due to the superposition of three factors: technical architecture optimization, model efficiency improvement, and scale effect. This trend will promote the further popularization of large model applications, which is in line with industry development laws.

---

[1] BOCOM International - 2024 Baidu Create AI Conference Minutes (https://pdf.dfcfw.com/pdf/H3_AP202404191630636523_1.pdf)
[2] 36Kr - In 2025, Are “Large Model Price Wars” No Longer Afraid of Losing Money? (https://m.36kr.com/p/3125094788382464)
[3] Wall Street News - Baidu Intelligent Cloud’s AI-related Revenue Grew by Nearly 300% in 2024 (https://wallstreetcn.com/articles/3741365)