Bytedance Seed 1.8 is ByteDance's generalized agentic model. It combines a Search Agent, Code Agent, and GUI Agent in one multimodal system with token-efficient visual encoding and three adaptive thinking modes.
import { streamText } from 'ai'
const result = streamText({ model: 'bytedance/seed-1.8', prompt: 'Why is the sky blue?'})Bytedance Seed 1.8 integrates three agent functions into one system. The Search Agent handles information retrieval across web and document sources. The Code Agent writes, debugs, and runs code. The GUI Agent interacts with graphical interfaces on desktop, web, and mobile using native vision rather than scripted automation, and operates software the way a human would.
Visual token efficiency is a core engineering focus. Bytedance Seed 1.8 reduces image encoding token requirements without sacrificing reasoning quality. This matters for GUI-heavy workloads where dozens of screenshots may pass through a single session. Three adaptive thinking modes calibrate processing depth to task complexity. They skip unnecessary compute on straightforward steps and use deeper reflection on ambiguous decisions.
In ByteDance's published benchmarks, Bytedance Seed 1.8 reaches 67.6 on BrowseComp-en, 87.8 on VideoMME (long-form video understanding), and 11.0 on ZeroBench (multimodal reasoning). It scores 62.0 on VLMsAreBiased and 47.2 on WorldTravel, up from its predecessor Seed 1.5-VL in those tables. Evaluations cover simulated workflows including travel planning, financial analysis, and software engineering. See https://docs.byteplus.com/en/docs/ModelArk/2123228 for methodology, tables, and comparisons.