Seed 1.8

Seed 1.8 is a ByteDance foundation model designed as a single, generalized agentic system that unifies perception, reasoning, and action. It accepts text and image input and pairs that multimodal understanding with native search, code execution, and graphical user interface navigation, allowing it to carry out end-to-end tasks such as retrieving information, driving a browser or app, running code, and chaining these steps into longer workflows. A large the cataloged API limit context window supports that iterative, multi-step style of work, while three adaptive thinking modes let the model shift its processing depth on the fly to balance careful reasoning against inference efficiency. The intended fit is real-world automation: applications that need autonomous decision-making and grounded reasoning across many tools and input types rather than a single-turn question-answering setup.

On reported benchmarks, Seed 1.8 is positioned as a strong agentic and visual reasoner. ByteDance reports a score of 67.6 on the BrowseComp-en web-browsing benchmark, where it claims the model surpasses other leading systems, and 87.8 on the VideoMME long-video understanding benchmark, signalling competence at sustained, multimodal comprehension alongside its tool-use strengths. Together, these results frame Seed 1.8 as a practical choice for builders who want one model that can read documents and video frames, plan, and act through GUIs and APIs, rather than assembling separate specialists for perception, reasoning, and tool orchestration.