What is Doubao Large Model 1.5
Doubao Large Model 1.5 is the latest large model launched by ByteDance. It adopts a large-scale sparse MoE architecture, equivalent to the performance of a Dense model with 7 times the activated parameters, and its overall score is superior toGPT-4oand models such as Claude 3.5 Sonnet on multiple evaluation benchmarks including knowledge, code, reasoning, and Chinese. Doubao Large Model 1.5 also introduced the Doubao Realtime Voice Model (Doubao-1.5-realtime-voice-pro) and the Doubao Vision Understanding Model (Doubao-1.5-vision-pro), featuring low-latency, interruptible voice conversation capabilities as well as stronger visual reasoning and document recognition capabilities. No data generated by any other model was used during the model training process.
Main features of Doubao Large Model 1.5
- Significantly enhanced overall capabilities: It performs at a globally leading level on multiple authoritative evaluation benchmarks such as knowledge (e.g., MMLU_PRO, GPQA), code (e.g., McEval, FullStackBench), reasoning (e.g., DROP), and Chinese (e.g., CMMLU, C-Eval), with overall scores superior to industry-leading models such as GPT-4o and Claude 3.5 Sonnet.
- Efficient model architecture and low cost: It adopts a large-scale sparse MoE architecture, equivalent to the performance of a Dense model with 7 times the activated parameters, far exceeding conventional efficiency in the industry. The self-developed server cluster solution supports low-cost chips, greatly reducing hardware costs.
- Comprehensive improvement in multimodal capabilities
- Doubao Vision Understanding Model (Doubao-1.5-vision-pro): It has been comprehensively upgraded in multimodal data synthesis, dynamic resolution, multimodal alignment, and hybrid training, with significantly enhanced capabilities in visual reasoning, text document recognition, and fine-grained information understanding.
- Doubao Realtime Voice Model (Doubao-1.5-realtime-voice-pro): It adopts a Speech2Speech end-to-end framework, supports end-to-end voice conversations, and features low latency and interruption at any time, and has been fully launched on the Doubao App.
- Deep thinking capability: Based on the Doubao 1.5 foundation model, through RL algorithm breakthroughs and engineering optimization, the deep thinking model Doubao-1.5-Pro-AS1-Preview was developed, showing leading performance in evaluations such as AIME.
- Data independence: No data generated by any other model was used during the model training process, and a completely autonomous data production system was built to ensure the independence and reliability of data sources.
Technical principles of Doubao Large Model 1.5
- Large-scale sparse MoE architecture: Doubao Large Model 1.5 adopts a large-scale sparse MoE (Mixture of Experts) architecture. Through pre-training with fewer activated parameters, it is equivalent to the performance of a Dense model with 7 times the activated parameters, far exceeding the conventional 3x leverage efficiency in the industry.
- Multimodal fusion technology: The model has been significantly upgraded in multimodal capabilities, supporting input and output of multiple modalities such as text, images, and speech.
- Efficient data processing and training: Doubao Large Model 1.5 did not use any data generated by other models during training. It relies on an independently built data production system, combining annotation teams and model self-play technology, to ensure the independence and reliability of data sources. Through a self-developed server cluster solution and optimization technology, the model significantly reduces hardware costs.
- Reinforcement learning and optimization framework: The Doubao Large Model team proposed the HybridFlow framework, a flexible and efficient reinforcement learning (RL) training framework that combines the advantages of single-controller and multi-controller approaches, significantly improving training throughput.
- Model Optimization and Inference Acceleration: Doubao Large Model 1.5 optimizes the model's inference efficiency through techniques such as fine-grained quantization and PD separation.
How to Use Doubao Large Model 1.5
- Doubao App: Doubao Large Model 1.5 has been rolled out in grayscale, and users can experience it in theDoubaoApp.
- Volcano Engine API: Developers can directly call the API throughVolcano Engine, supporting multi-scenario applications.
- Price Advantage: The original model price remains unchanged, with more added without increasing the price.
Project Address of Doubao Large Model 1.5
- Project Official Website：https://team.doubao.com/zh/special/doubao_1_5_pro
Application Scenarios of Doubao Large Model 1.5
- Sentiment Analysis and Feedback: Through sentiment analysis of speech and text, better understand user emotions and provide more targeted services.
- Intelligent Homework Tutoring: Help students solve problems in subjects such as mathematics and science, and provide problem-solving ideas and steps.
- Text Generation: Supports long-form text generation, suitable for news reporting, copywriting, story creation, etc.
- Video Generation: The Doubao video generation model can generate high-quality videos based on text or images, supporting dynamic poster and short video creation.
- Visual Understanding: The Doubao visual understanding model can identify objects and scenes in images and perform logical reasoning, suitable for problem analysis and chart analysis in the education field, etc.
- Multilingual Learning: Supports multilingual speech recognition and generation, and can be used for language learning and teaching.