Baichuan Intelligence

Baichuan Intelligence is a Chinese AI company that develops large language models (LLMs). 

The Baichuan model series represents a generational progression in language model development. The series began with Baichuan 1, followed by Baichuan 2, which includes Baichuan2-7B and Baichuan2-13B. These models, trained on 2.6 trillion tokens, show significant improvements over their predecessors across various benchmarks. For instance, Baichuan 2-7B achieves nearly 30% higher performance compared to Baichuan 1-7B on general benchmarks like MMLU, CMMLU, and C-Eval. Baichuan 4, released in 2022, offers a 10% overall improvement, along with a 14% boost in math and a 9% boost in coding performance over previous iterations.

Within the Baichuan series, several specialized models cater to specific tasks or domains. Baichuan2-192k, launched in October 2023, features an extended context window to manage around 350,000 Chinese characters in one instance, which is a unique characteristic, enhancing text understanding and summarization for long documents. Baichuan 3, introduced in January 2024, specializes in Chinese medical tasks and traditional Chinese literary composition. The company also mentions Baichuan2-53B, a closed-source model released in September 2023, which reportedly incorporates features like intention-understanding, information retrieval, and reinforcement learning, along with mathematical and logical reasoning abilities. 

Key customers and partnerships

In August 2024, Baichuan Intelligence and the National Children's Medical Center at Beijing Children's Hospital entered a strategic partnership to explore the integration of AI technology to improve both the quality and accessibility of pediatric healthcare resources.