ByteDance is understood to be in the early stages of pre-training a frontier model that could reach as many as 10 trillion parameters, a figure that would put it within touching distance of Anthropic’s Mythos 5, widely estimated by industry analysts to sit at around 8 trillion parameters.
For context, that would make ByteDance’s model more than three times larger than Moonshot AI’s Kimi K3, currently China’s biggest publicly released system at 2.8 trillion parameters. Pre-training alone typically takes three to six months, and the final parameter count won’t be locked in until later in the process.
Why the comparison to Mythos matters
Mythos was withheld from general release on capability grounds before Fable 5, its more constrained sibling, briefly reached the public. Within days, both were pulled under US export controls after officials raised concerns that Mythos had been jailbroken, a saga Capacity covered in detail at the time and again as access was gradually restored. The episode wasn’t just corporate drama. It fed directly into a live debate about how far national governments should go in gatekeeping frontier model access, a question that now looks set to resurface as ByteDance closes the gap.
Testing has also highlighted why Washington is so twitchy about Mythos-class systems in the first place. Democratic Senator Mark Warner told a Senate committee the model had reportedly identified weaknesses across nearly every classified government system it was tested against, saying it “broke into almost all of our classified systems, not in weeks but in hours.”
Anthropic has framed that capability as a defensive asset through its Project Glasswing initiative, arguing that “no one organisation can solve these cybersecurity problems alone.” Whether a ByteDance system with similar reach would be met with the same caution, or a very different one, is an open question worth watching.
The chip and capacity squeeze behind the headline
ByteDance flagged its intentions well before this latest report, committing around $23 billion to AI capital investment in 2026, with roughly half earmarked for semiconductors. Training a model at this scale demands enormous, coordinated GPU clusters, exactly the kind of infrastructure that US export restrictions have made harder for Chinese firms to source.
That has pushed ByteDance down several parallel tracks at once. It has been linked to a custom CPU programme to reduce reliance on external suppliers, and reportedly held early talks with Qualcomm on chip design services. Meanwhile, Nvidia’s own route into ByteDance has been complicated by unresolved Know Your Customer conditions attached to any H200 licence, leaving the deal in limbo. Domestic silicon has filled some of the gap: ByteDance and Alibaba have reportedly placed fresh orders with Huawei for its Ascend chips, a shift Omdia’s He Hui described by noting that “Huawei’s Ascend chips are the country’s best homegrown alternative to Nvidia.” That pattern accelerated further after DeepSeek optimised its V4 model specifically for Ascend hardware, triggering a wider procurement scramble among Chinese hyperscalers.
Put together, a ByteDance model approaching Mythos scale isn’t just a research milestone. It’s a signal of sustained, multi-year demand for GPU capacity, power and cooling infrastructure inside China, delivered through whatever mix of domestic and grey-market hardware companies can secure.
What it means for future deals
Three things stand out in regard to future deals. First, the compute arms race is broadening rather than narrowing. Chinese labs including Moonshot, DeepSeek and MiniMax have already faced accusations from Anthropic of extracting Claude’s capabilities through unauthorised distillation, illustrating how quickly capability gaps can close even without matching hardware access. A ByteDance model built independently, at genuine frontier scale, would be a different order of challenge and would likely prompt renewed scrutiny of chip export enforcement.
Second, expect the hyperscale infrastructure conversation to keep tilting towards sovereignty and control. As one Capacity source put it during coverage of the Fable 5 fallout, “policy intent must be matched by tangible outcomes.” Operators and investors weighing exposure to Chinese AI infrastructure, or to the semiconductor supply chains feeding it, will be watching how Washington and Beijing respond to this latest escalation before committing further capital.
Third, there’s genuine white space for coverage of what comes next on the enterprise side. Western frontier labs are already moving to lock in commercial partnerships around their most capable, most restricted models, seen recently in Anthropic’s tie-up with TCS to bring enterprise-grade Claude deployments to regulated industries. If ByteDance succeeds in fielding a genuine Mythos rival, similar enterprise and government partnership models could emerge on the Chinese side, particularly across Southeast Asia, the Gulf and parts of Africa where data sovereignty concerns already favour non-US providers. There’s also an open story in how colocation and cloud providers outside China position themselves: as a neutral option for enterprises wary of both US export risk and Chinese state proximity, or as a bridge for firms wanting access to whichever frontier model proves cheapest to run at scale.
Anthropic itself has acknowledged the sensitivity of what these models are now capable of. Following incidents in which its own Claude models were used to breach three companies during testing, the firm has been open about the dual-use risk built into agentic, cyber-capable AI. If ByteDance’s model reaches anywhere near the scale being reported, the industry will need to have the same conversation about a Chinese-built system, and data centre operators, chip suppliers and policymakers will all have a stake in how that plays out.
For now, ByteDance’s model remains in pre-training, with its final size still to be confirmed and a public release, if it comes at all, likely still months away. But the intent behind the project is unmistakable, and the infrastructure implications are already being felt.
Related stories
Huawei launches new chip framework to turbocharge China market leadership
Nvidia, ByteDance AI chip deal hinges on Trump-era conditions
