Youdao Says AI Competition Is Entering a 'Model + Agent + Workflow' Era as It Shows Agent Products in Beijing

CEO Zhou Feng said the real change of the past year is that AI is no longer a tool users must seek out and invoke, but an assistant embedded in existing workflows, and that work methods should even be designed around agents. In his description, the model determines how smart AI is and sets the capability ceiling, the agent makes it capable by turning model output into action, and the workflow delivers reliable output in real business settings.

Youdao's system places a Ziyue model matrix at the bottom for multimodal reasoning and real-time voice interaction, an agent product layer in the middle, and office, learning and marketing scenarios on top. In advertising, Youdao built agents across the chain of creative generation, influencer selection and placement optimization. iMagicBox generates digital-human videos and localized marketing assets from one image and one instruction; the company said this cuts material production costs by 80 percent and lifts click-through rates by 30 percent. AI-generated materials rose twelvefold over two years to 25 percent of platform assets. YODA draws on more than 2,000 real project records for influencer insight and marketing decisions, and Xiaozhi 2.0 automates data aggregation and account management to speed up cross-account queries and data synchronization.

Chief scientist Duan Yitao presented the Ziyue model matrix and a new Ziyue Live simultaneous interpretation model. Youdao upgraded Ziyue to version 4.0 in May, releasing multimodal reasoning, cross-lingual zero-shot and translation models; the first two were open-sourced, and Duan said they are at the frontier among models of comparable size. Youdao also open-sourced a Ziyue streaming ASR model aimed at voice-agent-grade speech recognition, supporting real-time streaming input and output and full-duplex interaction rather than turn-taking. Duan introduced Youdao Youju, a human-AI collaboration agent that draws on papers, patents and legal provisions, attaches checkable sources to each argument, and can write papers or generate slides. He argued that a valuable AI product should be a workbench plus a live assistant rather than a chat window that merely finishes tasks, calling Youju a harness for both the model and the human brain.

Wang Ning, who leads LobsterAI, released version 2.0, compatible with the OpenClaw 2.0 ecosystem, with new Sites and Teams features, one-click public URLs carrying share codes for task outputs, and usage-based billing instead of per-seat fees. He framed the next stage of office agents around organization, delivery and trust, saying an office agent should amplify a whole team rather than serve one person, that output must be easy to send onward, and that the competition is over which agent is worth entrusting. He said the team recently published what it described as China's first user behavior report on office agents, and that a mobile product is coming soon.

Zhang Yi, head of the smart applications unit, said voice will be a major entry point as AI moves from making people adapt to technology toward adapting to people. Youdao's AI Interpreter, launched in March as what the company called the industry's first simultaneous interpretation agent, has served more than 25 million users, with usage nearly doubling year on year. Bage Shuo, described as NetEase's first AI-native voice agent, converts natural speech into structured text meant to be reused directly, addressing the gap between typing speed and thinking speed. Hi Echo has more than 10 million cumulative users and uses end-to-end full-duplex voice dialogue with a self-developed goal-driven self-evolving learning system, Echo HELIX, for an oral English tutor agent that is low-latency, interruptible and able to plan personalized tasks. Chris Mathews, managing director for English and assessment at the British Council, attended and spoke about Youdao's collaboration with IELTS, and a child at the event interacted with a Qiyuan robot running Hi Echo.

Gao Huitruan, head of the smart hardware unit, said AI's value lies in extending people's abilities and reducing their load rather than replacing them. The OpenPods AI earbuds link recording, transcription, summarization and translation, covering listening, note-taking and follow-up action in meetings. The product supports real-time audio and video translation in more than 20 languages and uses voice cloning to retain a speaker's timbre; the charging case can record, translate and play back without a phone. OpenPods separates speakers, cleans up repetition and restatements in speech, converts recordings into written minutes and to-dos, and lets users ask questions about notes with sources cited. To-dos can then be turned into executable task lists, moving the earbuds from recording assistance toward carrying out tasks.

Editor's Summary

NetEase Youdao used its Sept 16 Beijing AI Open Day to argue that AI competition is shifting from models alone to systems combining models, agents and workflows, and to release or update a broad set of products including LobsterAI 2.0, the Ziyue Live interpretation model and OpenPods. The announcements center on embedding agents into office, learning, marketing and cross-language tasks rather than on standalone model benchmarks.