Mr Viral
On December 30th, ByteDance‘s Doubao large model released the technological advancements across all fields for 2024. It claimed that since its debut on May 15th, within seven months, the model’s capabilities in general language, video generation, voice dialogue, visual understanding and other aspects have reached the international forefront.
As introduced by them, as of December 2024, the latest version of Doubao’s general model “Doubao-pro-1215” has improved its comprehensive abilities by 32% compared to May. It is now fully aligned with GPT-4o and performs even better in complex scenarios such as mathematics and professional knowledge tasks. Additionally, its inference service price is only one-eighth of GPT-4o. Through massive data optimization and innovative model architecture including improving sparsity and introducing reinforcement learning techniques among others, this model has significantly enhanced precision in understanding and quality in generation.
Doubao also disclosed for the first time its long-text capability with a window size of three million words for large models. According to their introduction, it can easily read “hundreds” of academic reports at once with a processing delay of only 15 seconds per million tokens. Behind this capability are algorithms like STRING which relate contextually to data; these greatly enhance LLM’s ability to utilize vast external knowledge sources while sparse distribution solutions reduce latency down to tens-of-a-second level.
SEE ALSO: ByteDance’s Doubao Vision Understanding Model is 85% Cheaper than the Industry Average Price
Sign up today for 5 free articles monthly!
GIPHY App Key not set. Please check settings