ByteDance plans to release an AI assistant similar to Meta's Muse in the coming weeks, a move meant to shore up its position in AI video, Bloomberg reported, citing people familiar with the matter. The information comes from a closed-door all-hands meeting, where co-founder Zhang Yiming told the AI team that ByteDance has the ability to compete with Anthropic and OpenAI at the frontier, but not until 2029 or 2030 at the earliest.
The report does not directly quote Zhang. Everything attributed to him was relayed by sources who asked not to be named. A ByteDance spokesperson did not respond to Bloomberg's emailed request for comment, so none of this can be independently verified and it should be read as the report's account.
What else was said at the meeting
According to the report, ByteDance is prepared to invest up to $70 billion to turn Seedance into China's leading video AI platform. Zhang acknowledged at the meeting that the datasets used to pretrain the Seed team's large language models are far smaller than those of frontier rivals. He also reiterated that ByteDance should not rely on model distillation and should take the long view.
Meta's Muse launched in September this year and runs on Meta's own hardware, including its smart glasses and an upcoming device called Muse Charm. Whether ByteDance's assistant will be paired with hardware is, per the report, still unclear.
The line about not relying on distillation carries particular weight in a Chinese context. Using a large model's outputs to train a smaller one is the fastest way for a follower to close the gap, and it is also a practice that overseas labs have publicly criticized. Going by the account relayed in the report, skipping that shortcut and putting the catch-up date three or four years out are consistent with each other. What was actually said in the meeting, and whether there is a more specific roadmap, rests solely on anonymous sources; ByteDance has not confirmed anything publicly.
Key moments in Seed's year
Strung together, the dates mentioned in the report show where ByteDance's large model team has stood this year.
In June, Seed shipped its most recent major update, Seed 2.1. The team itself wrote at the time:
"There remains room for improvement in the most challenging open-ended tasks and cutting-edge research problem-solving"
In August, the Financial Times reported that ByteDance was training a model with as many as 10 trillion parameters, comparable in scale to Anthropic's Claude Mythos. In October, Seed set up a standalone safety and alignment department and recruited about 10 core researchers.
The team numbers roughly 1,000 researchers and engineers. Its head, Wu Yonghui, was brought back by Zhang Yiming from Google DeepMind in 2025, and ByteDance has also offered Seed employees favorable equity arrangements. Bloomberg's assessment is that among China's major AI players, Seed is the only one that has neither produced a large language model ranking in the top tier of global benchmarks nor opened up its model weights to developers.
What users in China might see
For users in China, ByteDance's AI entry points are already quite consolidated. Doubao is its main consumer product, TRAE and Coze have previously been folded into the Doubao family, and Seedance and Seedream cover video and image generation respectively. If a Muse-like agent does launch, the most likely form is one more assistant inside this lineup that can run errands for the user, rather than a brand-new standalone app, though that is only a guess inferred from the existing setup.
Meta's Muse is not currently available in mainland China, so users there can hardly compare the two directly. Once ByteDance's product actually appears in a few weeks, the first two questions for Chinese users will be whether it can call Seedance to generate video and whether it can operate services inside Douyin and Toutiao.
The $70 billion figure also needs context: it is described as an upper limit on investment. Over how many years it would be spent and whether it includes compute purchases are not broken out in the report.
Sources: Bloomberg (for the all-hands meeting content, the $70 billion investment ceiling and Seed team size), The Economic Times report on the same story, CocoLoop, and the Financial Times' earlier reporting on the scale of Seed's model training.