ByteDance is reportedly weeks away from its own Muse-style AI agent, and Zhang Yiming told staff catching OpenAI will take until 2029

The backstory

Muse is the AI assistant Meta launched in September. It is not just a chat box. It talks with you by voice, shows up as an animated character, and helps with everyday tasks — what the industry calls an AI agent, meaning an AI that does things on your behalf instead of only answering questions. Meta is putting it on its smart glasses and is preparing a small dedicated gadget, the Muse Charm. Bloomberg describes it as a big hit for Meta.

ByteDance owns TikTok, its Chinese sibling Douyin, and Doubao, one of China's most popular chatbots. It is strong in AI video: its Seedance models became a top-tier, money-making video generator, filling the gap left when OpenAI shut down Sora. But on the core technology, large language models, ByteDance has not yet shipped a model near the top of global leaderboards, and it is the only major Chinese player that hasn't released open model weights.

What was reported

According to Bloomberg, ByteDance plans to release an assistant that plays a role similar to Muse within weeks, hoping to build on its lead in video and close the gap with top U.S. labs in as little as three years. The product's name, features, and whether it comes with hardware are not known.

The more unusual part is what Zhang Yiming reportedly said. Zhang, who rarely appears in public but now spends much of his time shaping ByteDance's AI strategy from behind the scenes, told the AI team at a recent all-hands meeting that ByteDance can compete with Anthropic and OpenAI at the cutting edge — but not before 2029 or 2030. He said the datasets used to pretrain Seed, the company's in-house models, are orders of magnitude smaller than those of frontier rivals. Pretraining is the stage where a model reads enormous amounts of text to learn the basics; "orders of magnitude" means a gap of ten times, a hundred times, or more.

He also reiterated that ByteDance should not rely on distillation, the practice of training a model on the outputs of a stronger model. Think of it as copying the top student's homework: it gets you closer fast, but rarely gets you ahead. White House officials and Anthropic CEO Dario Amodei have criticized Chinese labs for doing it.

Then there's the money. Bloomberg reported that ByteDance plans to spend as much as $70 billion to turn Seedance into China's leading video AI platform. Separately, the Financial Times reported in August that ByteDance was training a model with as many as 10 trillion parameters.

To be clear about sourcing: these details come from people familiar with a private meeting, and ByteDance did not respond to Bloomberg's request for comment. This is a press report, not an official announcement. What ByteDance actually ships, and when, will depend on the company.

Why it matters

First, it is rare for a Chinese tech leader to put such a cautious timeline on catching up. Zhang's "2029 at the earliest" amounts to admitting that the top U.S. labs are still years ahead on frontier models.

Second, it shows the strategy. Instead of shortcuts, ByteDance is betting on more data and compute, while using its strongest assets — video AI and a massive user base — to push an agent product now. That echoes Meta's approach: the model doesn't have to be the world's best if the assistant lives inside apps people already open every day.

What it means for you

If the assistant launches on schedule, it will most likely show up inside apps like Doubao, Douyin, or TikTok-adjacent products — a talking AI helper that can plan trips, edit videos, or look things up. That's convenient, but agents work best when they can see more of your personal data, like messages, location, and calendars, so the privacy settings will be worth a careful look.

Longer term, a company willing to commit up to $70 billion means competition in AI video and assistants will heat up, which usually brings ordinary users more tools at lower prices.

Original sources published: Bloomberg (@business on X, linking its report), Nov 11, 2026, 18:35 Beijing time (UTC+8); The Economic Times (Bloomberg syndication), Nov 11, 2026, 18:48 Beijing time; ifeng Tech, Nov 11, 2026, 18:59 Beijing time