Doubao 2.1 Pro Model Update Enhances Agent and Coding

ByteDance released Doubao 2.1 Pro (0915 version) with improvements in agent task delivery, multimodal coding, multimodal understanding, and inference cost. The model can read design drafts and screenshots to generate frontend code, and reduced image/video token consumption by 30%.
Key insights
Agent and Coding Improvements
The model strengthens evidence tracing and data verification in multi-step tool calls. In a financial research example, it dispatched 500+ sub-agents, retrieved 1000+ web pages, and cross-referenced maritime tracks and satellite images. For multimodal coding, it can read design drafts, screenshots, and even operation recordings to generate frontend code. It understood 280,000 lines of Java code to recreate a mobile page for a legacy ERP system.
Cost Reduction and Availability
Image and video inference token consumption is reduced by over 30% compared to the previous generation. The API has two endpoints: Doubao-Seed-2.1-pro-0915 for version locking, and Doubao-Seed-Evolving for automatic updates. The model is integrated with Doubao Work and TRAE.