ByteDance is preparing to launch a Muse-like AI assistant in the coming weeks, according to people familiar with a recent all-hands meeting of the company’s AI team. The move is designed to build on ByteDance’s strength in video artificial intelligence and help close the gap with leading American labs in as little as three years.
Co-founder Zhang Yiming addressed researchers, engineers and support staff at the meeting, laying out the plan and the priorities that will guide the project. He believes ByteDance can eventually compete with Anthropic and OpenAI at the frontier of AI development, though he cautioned that reaching that level would take until at least 2029 or 2030.
A product in the mould of Meta’s Muse
The planned agent is expected to serve a similar role to Meta’s Muse, which the US social media giant introduced in September. Muse combines voice interaction with an animated avatar and is being rolled out across Meta’s smart glasses and an upcoming dedicated device called the Muse Charm.
It remains unclear whether ByteDance intends to pair its own agent with similar hardware. What is clear is that the company sees a consumer-facing agent as a way to turn its AI research into a product that can compete directly with the most advanced offerings from Silicon Valley.
Data and the distillation debate
Top of mind for Zhang is the quality and scale of training data. He told staff that ByteDance’s Seed large language models are pre-trained on datasets orders of magnitude smaller than those used by frontier rivals. That gap in data is one of the main obstacles standing between ByteDance and the cutting edge.
Zhang also reiterated his view that ByteDance should avoid distillation, the practice of training smaller models on outputs from larger ones. The technique has drawn criticism from White House officials and Anthropic chief executive Dario Amodei, who have accused some Chinese labs of using distillation to build on top of American models.
The Seed unit’s long game
ByteDance’s Seed division, which now employs roughly 1,000 researchers and engineers, is central to the effort. The unit is led by Wu Yonghui, a former Google DeepMind researcher whom Zhang recruited in 2025. Seed’s last major language model release was Seed 2.1 in June, and the unit has not yet produced a model that ranks among the very best on global benchmarks or made its weights available to outside developers.
In October, Seed became the first major Chinese AI lab to set up a standalone safety and alignment team, a step that brings it closer to the practices of leading Western labs.
Video as the battleground
Where ByteDance currently leads is in video generation. Its Seedance models have filled the space left after OpenAI shut down Sora, turning ByteDance's vast pool of video content into a top-tier generation and monetisation platform. The company is reportedly prepared to spend as much as $70 billion on AI infrastructure to press that advantage.
Beyond video, ByteDance is expanding its AI footprint outside China. The company secured a $29.6 billion loan in September from close to three dozen banks, with proceeds earmarked for AI initiatives and data centres under development in Southeast Asia.
What comes next
A ByteDance spokesperson did not respond to a request for comment. If the timeline Zhang laid out holds, the launch of a ByteDance agent would mark one of the most significant entries yet by a Chinese company into the emerging market for AI assistants that combine conversation, avatar interfaces and task completion.
The coming weeks will show whether ByteDance can translate its video AI lead into a broader consumer product, and whether its long-term bet on original research can close the distance with OpenAI and Anthropic faster than Zhang's cautious timeline suggests.
