On January 22, 2025, ByteDance's Volcano Engine officially released the Doubao Model 1.5 and fully launched it on the Volcano Ark platform. The Doubao Model 1.5 has achieved significant performance improvements in various fields, reaching a globally leading level of overall capability, marking another important breakthrough for ByteDance in the field of artificial intelligence.
The Doubao Model 1.5 includes several versions, among which Doubao-1.5-pro has achieved the best scores on multiple authoritative evaluation benchmarks in knowledge, coding, reasoning, and Chinese, outperforming top industry models such as GPT-4o and Claude 3.5 Sonnet. Doubao-1.5-lite, on the other hand, excels among lightweight language models, its performance even rivaling the previous Doubao-pro-32k-0828 version, providing users with a better cost-performance ratio. Additionally, Doubao-1.5-vision-pro has undergone comprehensive upgrades in multi-modal data synthesis, dynamic resolution, and multi-modal alignment, enhancing visual reasoning and fine-grained information understanding, achieving leading performance on several authoritative evaluation benchmarks.
The release of Doubao Model 1.5 also introduced the Doubao Real-Time Voice Model, enabling end-to-end voice conversations with low latency and the ability to interrupt during dialogue, bringing new breakthroughs to the field of voice interaction. Volcano Engine plans to launch corresponding API services through the Ark platform in the first half of the year, further promoting the widespread application of voice technology.
In terms of technical architecture, Doubao Model 1.5 adopts a large-scale sparse MoE architecture, achieving the performance of a dense model with the equivalent of 7 times the activation parameters through a smaller number of active parameters, far exceeding conventional industry efficiency. Meanwhile, ByteDance's self-developed server cluster solutions and network card technology significantly reduce hardware costs, optimize small packet communication efficiency, and ensure the stability and efficiency of multi-machine distributed inference. Furthermore, Doubao Model 1.5 did not use any data generated by other models during its training process, establishing a completely independent data production system that ensures the independence and reliability of data sources.
It is noteworthy that despite the significant improvements in performance and functionality, the price of Doubao Model 1.5 remains unchanged, adhering to the principle of "more features at no extra cost," aiming to promote the accessibility of AI technology, allowing more enterprises and developers to benefit from this advanced technological achievement.
