ByteDance has launched a new beanbag model 1.5Pro, surpassing GPT-4o and Claude3.5Sonnet in multiple benchmark tests, marking its significant progress in the field of artificial intelligence. This model uses an innovative sparse MoE architecture to achieve performance equivalent to the 7-times parameter Dense model with fewer activation parameters, and the efficiency is increased by about 3 times. In addition to the core model upgrade, Doubao visual understanding model and real-time speech model were simultaneously released, further enhancing multi-modal processing capabilities and voice interaction experience.
ByteDance officially launched its latest Doubao model 1.5Pro (Doubao-1.5-pro). This new model performs well in comprehensive capabilities in multiple fields, successfully surpassing the well-known GPT-4o and Claude3.5Sonnet in the industry. . The release of this model marks another important step forward for ByteDance in the field of artificial intelligence.
Doubao 1.5Pro adopts a new sparse MoE (Mixed Expert) architecture and uses smaller activation parameters for pre-training. The innovation of this design is that it can provide Dense model performance equivalent to 7 times the activation parameters, making it far more efficient than the industry's conventional MoE architecture, bringing about a 3-fold efficiency improvement. This design makes the Doubao model score even better on multiple evaluation benchmarks such as knowledge, code, reasoning, and Chinese.
In addition to the upgrade of the main model, ByteDance also released Doubao visual understanding model Doubao-1.5-vision-pro and Doubao real-time voice model Doubao-1.5-realtime-voice-pro. The new visual understanding model has undergone comprehensive technical upgrades in multi-modal data processing, dynamic resolution and fine-grained information understanding, further improving its capabilities in visual reasoning and text understanding. At the same time, the launch of the real-time speech model enables Doubao App to achieve a smoother voice conversation experience, with low latency and the ability to interrupt at any time during the conversation.
ByteDance officially stated that the Doubao model did not use any data generated by external models during the training process, ensuring the independence and reliability of the model. In addition, the pricing of all new products will remain unchanged, and users can directly experience new features in the Doubao App.
This conference not only demonstrated ByteDance's continuous innovation capabilities in the field of AI, but also provided developers with strong API support, further promoting the popularization and application of artificial intelligence technology.
The launch of the Doubao large model 1.5Pro, as well as the supporting visual and voice models, demonstrate ByteDance’s strong strength and technological innovation in the field of AI. Its high efficiency, high performance and emphasis on user experience indicate that AI applications will be more convenient and powerful in the future.