ByteDance's Doubao large model has released its 2024 Technical Progress Report today, revealing that its latest version, Doubao-pro-1215, has achieved comprehensive alignment with GPT-4 in overall performance and demonstrates stronger capabilities in certain specialized fields. This advancement marks the official entry of Chinese large model technology into the global top tier.
Since its debut in May this year, the Doubao large model has achieved a 32% improvement in capabilities in just seven months. According to official sources, significant progress has been made in understanding accuracy and generation quality through the optimization of massive data processing and innovative model architecture, including increased model sparsity and the introduction of reinforcement learning techniques. Notably, in complex scenarios such as mathematics and specialized knowledge, its performance even surpasses that of GPT-4, while the service price is only one-eighth of that of the latter.
It is worth noting that Doubao has publicly unveiled its capability to process ultra-long texts of 3 million characters, which means it can handle content equivalent to "hundreds of academic reports" simultaneously. By employing context-related data algorithms such as STRING and optimized sparsification and distribution schemes, Doubao has managed to keep the processing delay for millions of tokens within 15 seconds, significantly enhancing the model's efficiency in processing vast amounts of external knowledge.
This technological breakthrough not only showcases the rapid development of AI technology in China but also suggests that the widespread adoption of large models may accelerate due to better cost-effectiveness.
