ByteDanceUnderBean curdThe Big Model announced its 2024 technology progress across the board on December 30, claiming that in the seven months since its debut on May 15, the model's capabilities in general language, video generation, voice dialog, and visual understanding have crossed into the top tier of the international arena.
It is reported that as of December 2024, the latest version of the Doubao generic model Doubao-pro-1215 has increased its comprehensive capacity by 32% compared to May, and it has beenFull alignment GPT-4oThe model has been optimized to perform even better in some complex scenarios such as math and professional knowledge tasks, while its reasoning service price is only one-eighth of GPT-4o. It has gone through massive data optimization and model architecture innovations, including improving model sparsity and introducing reinforcement learning, which has led to a significant increase in the model's comprehension accuracy and generation quality.
Beanbag also revealed its big model to the public for the first time Long text in 3 million word windowAbility. According to the introduction, it can easily read "hundreds" of academic reports at a time, and the processing delay is only 15 seconds per million tokens. According to 1AI, behind it is STRING and other contextualized data algorithms, which can significantly improve LLM The ability to leverage massive external knowledge, sparsification and distributed schemes reduce latency to ten seconds.