Seed 1.8 is ByteDance's generalized agentic foundation model that unifies perception, reasoning, and action in a single system. It accepts text and image inputs and folds native search, code, and graphical user interface agent capabilities into the same backbone, enabling end-to-end task automation across information retrieval, GUI navigation, code execution, and multi-step workflows. Three adaptive thinking modes let the model shift its processing style with task complexity, trading reasoning depth against inference efficiency rather than forcing a single fixed behavior. Together these choices point to a model aimed at applications that need autonomous decision-making and grounded reasoning across multiple modalities and tool types, rather than a pure chat or single-modality model.
On agentic and multimodal evaluation, the model reports a BrowseComp-en score of 67.6, a web-browsing benchmark where ByteDance describes it as outperforming other leading systems, and a VideoMME long-video understanding score of 87.8, signaling competence beyond pure text reasoning. The large that quick-info value window supports long, multi-step workflows where the model has to track evidence, code, and tool outputs together, which fits use cases like research assistants, automated browser workflows, and code-driven pipelines. The combination of vision input, tool use, structured output, and configurable reasoning depth makes Seed 1.8 a practical fit for teams building agent stacks that need a single model to perceive inputs, plan over many steps, and act through external tools.
