In a packed meeting room in Beijing, the man who built TikTok stood in front of the engineers who are supposed to catch up with America — and told them the truth first.
ByteDance can compete with OpenAI and Anthropic at the cutting edge, co-founder Zhang Yiming said. But not before 2029 or 2030.
Then came the plan for right now: a new AI assistant, modelled on Meta’s Muse, arriving within weeks.
According to Bloomberg, it is ByteDance’s most direct attempt yet to close a gap that has defined the AI race for two years — the gap between China’s biggest tech companies and the American labs setting the pace.
What ByteDance’s Muse-like AI assistant is supposed to do
The new assistant will serve a similar role to Muse, the voice-based agent Meta introduced in September, according to people familiar with the all-hands meeting who spoke to Bloomberg.
Muse is not a chatbot window. It is a voice-first companion with an animated avatar, built to help with everyday tasks, and Meta has been pushing it into its hardware — its smart glasses and an upcoming gadget called the Muse Charm, made just for the agent.
ByteDance’s version is expected in the coming weeks, the report said. Few product details are public yet. It is not clear what the assistant will do first, whether it will be offered outside China, or whether ByteDance will build companion hardware of its own.
What is clear is the strategy behind it. The assistant is meant to build on ByteDance’s existing strength in AI video, where its Seedance models are already a paid product, and turn that strength into something users talk to every day.
Muse set the template ByteDance is chasing
Meta launched Muse in September, and by Meta’s own telling it has been a hit. More than 10 million people in the United States and Canada set up Muse in the six weeks after launch, with around a quarter using it daily, according to PYMNTS.
Meta chief executive Mark Zuckerberg used the company’s third-quarter earnings call to hold Muse up as proof that its AI strategy is working. The assistant has become what PYMNTS called “a new centerpiece” for Meta’s AI ambitions.
That success is exactly what makes it a template. A voice agent with a face, living on glasses and in pockets, is a very different product from a chatbot you visit on a website. It is personal, always on, and hard to copy — which is why ByteDance wants one of its own.
Meta has already shown the playbook works with its own cheaper models too. Its tiny Muse Ember model, launched this month, matches far bigger rivals on coding tasks at a fraction of the cost — a reminder that the assistant race is now about efficiency as much as raw power.
The data problem holding ByteDance back
Zhang’s message to his team was not a victory speech. It was a diagnosis, and the diagnosis was data.
ByteDance’s Seed language models are pre-trained on datasets orders of magnitude smaller than those used by its frontier competitors, according to the people who spoke to Bloomberg.
Zhang told the team the company should not rely on distillation — the shortcut of training smaller models on the outputs of bigger ones — to make up the difference.
The Seed team numbers around 1,000 researchers and engineers, led by Wu Yonghui, who joined ByteDance from Google DeepMind in 2025. Its last major language-model update, Seed 2.1, arrived in June.
Meanwhile, the American labs have kept moving. Anthropic has launched its Claude office suite and filed for one of the largest IPOs in history. OpenAI has put ChatGPT into businesses, workplaces and the AWS Marketplace.
ByteDance has Doubao, China’s most popular ChatGPT-style bot, but Seed has neither released a model at the top of global benchmarks nor opened its weights to developers.
Video first, then the assistant
If ByteDance is behind on language models, it is ahead somewhere else — and the assistant plan leans on that.
The company’s Seedance video models are a genuine commercial success, sold as a paid product while rivals are still figuring out how to charge for video generation.
Video is the one place where ByteDance can credibly claim to lead, and the new assistant is designed to grow out of that lead rather than pretend it does not exist.
That matters because an AI assistant is only as good as what it can do for you.
A text chatbot answers questions. An assistant that understands video — that can make it, edit it, search it, explain it — can do things its rivals cannot. ByteDance’s bet is that video is the feature that makes its assistant worth switching to.
It is the same logic Chinese rivals have followed. Alibaba’s Qwen4-Max, previewed this month, leads with multimodal reasoning across text, image and video.
Meituan’s LongCat-2.5 opened its weights to anyone willing to download them. The Chinese labs are not copying the American playbook page by page — they are finding the pages where they can win.
Why 2029 is the real deadline
The most striking part of Zhang’s message was not the assistant. It was the timeline.
Three years is a long time in AI. Models that lead today can look dated in six months, and the frontier moves faster every year. Telling a thousand engineers that the catch-up will take until 2029 or 2030 is an admission that the gap is structural, not a single missing feature.
It is also a warning to anyone expecting a quick fix. Data at frontier scale cannot be bought overnight, and the compute to train on it is expensive — reported plans alongside the assistant suggest ByteDance could spend up to $70 billion on the effort, a figure the company has not confirmed publicly.
But the timeline cuts both ways. If ByteDance really does close the gap by 2029, the AI race stops being a two-country contest between a handful of American labs and becomes something much bigger — with the company behind TikTok, the world’s most addictive app, in the fight.
What happens next
ByteDance did not respond to Bloomberg’s request for comment, and the company has not announced the assistant publicly. The first real test will be the launch itself: what the assistant can actually do, how it sounds, and whether people outside China ever get to try it.
The second test will be quieter. Watch whether Seed’s next model update closes the benchmark gap, whether the assistant shows up on any hardware, and whether Doubao’s users start spending their time with the new assistant instead.
For now, the message from Beijing is clear enough. ByteDance knows exactly how far behind it is, it knows why, and it has a plan — starting with an assistant that looks a lot like the one Meta built. The race to build the AI that lives in your pocket just got a new runner.
Source: PYMNTS, reporting on Bloomberg.
