字节跳动豆包大模型2.0发布，多数基准达SOTA水平

Core Insights - ByteDance announced the official launch of Doubao 2.0, which has undergone systematic optimization for large-scale production environments, enhancing its capabilities in efficient reasoning, multimodal understanding, and complex instruction execution [1] Model Features - Doubao 2.0 includes three general agent models: Pro, Lite, and Mini, as well as a Code model, designed to adapt flexibly to various business scenarios [1] - Doubao 2.0 Pro is now available on the Doubao App, desktop, and web versions, allowing users to experience the "expert" mode for interactive dialogue [1] Performance Enhancements - Doubao 2.0 has significantly upgraded its multimodal capabilities, achieving state-of-the-art (SOTA) levels in various visual understanding tasks, with Doubao 2.0 Pro scoring highest in most relevant benchmark tests [2] - The model has improved its understanding of time series and motion perception, leading in key assessments like TVBench and surpassing human scores in the EgoTempo benchmark [4] Long-Range Task Execution - Doubao 2.0 Pro has enhanced long-range task execution capabilities, outperforming GPT 5.2 in SuperGPQA and achieving first place in HealthBench, with overall performance comparable to Gemini 3 Pro and GPT 5.2 in scientific domains [5] - In reasoning and agent capability evaluations, Doubao 2.0 Pro achieved gold medal results in IMO, CMO math competitions, and ICPC programming contests, demonstrating strong mathematical and reasoning skills [5] Cost Efficiency - Doubao 2.0 has reduced inference costs significantly, with model performance comparable to top industry models while lowering token pricing by approximately an order of magnitude [8] Code Model Features - The Doubao 2.0 Code model is optimized for programming scenarios, enhancing code library interpretation and application generation capabilities, and has been integrated into TRAE for improved functionality [9] - An example project, "TRAE Spring Festival Town · Year of the Horse Temple Fair," illustrates the model's ability to construct complex applications efficiently with minimal prompts [9]