The beanbao model claims to have equaled GPT-4 and disclosed for the first time its 3 million long text capabilities.

Doubao Big Model, a subsidiary of Bytedance, released its 2024 annual technology progress report today, revealing that its latest version, Doubao-pro-1215, has achieved full alignment with GPT-4 in terms of overall performance, and has shown stronger capabilities in some professional fields. . This progress marks that China's large model technology has officially entered the first echelon in the world.

Since its debut in May this year, the large bean bag model has achieved a 32% capacity improvement in just 7 months. According to the official introduction, Doubao has made significant progress in understanding accuracy and generation quality by optimizing massive data processing and innovating model architecture, including improving model sparsity and introducing reinforcement learning and other technical means. Especially in complex scenarios such as mathematics and professional knowledge, its performance even surpasses GPT-4, while the service price is only one-eighth of the latter.

It is worth noting that Doubao disclosed for the first time its ultra-long text processing capability of 3 million words, which means that it can simultaneously process the content equivalent to "hundreds" of academic reports. By using contextual data algorithms such as STRING, as well as optimized sparsification and distribution solutions, Doubao controls the processing delay of millions of tokens within 15 seconds, greatly improving the model's processing efficiency for massive external knowledge.

This technological breakthrough not only demonstrates the rapid development of China's AI technology, but also indicates that the popularization of large model applications may be accelerated due to better cost performance.