ByteDance Is Building One of the Biggest AI Models Ever Made

Key points

• ByteDance is training an AI model with as many as 10 trillion parameters, according to three people with knowledge of the project.
• That figure is three times larger than Kimi K3, the biggest Chinese AI model released publicly to date.
• The model is in pre-training, a phase that typically runs three to six months before any public release.
• The final parameter count hasn't been locked in and will only be confirmed at a later stage.
• If completed, the model would approach the scale of Anthropic's most advanced AI systems.

ByteDance, the Chinese company behind TikTok, is quietly building what could become one of the largest AI models in the world. Three people familiar with the project told Ars Technica the model may reach 10 trillion parameters, a number worth unpacking.

A parameter is a tiny internal setting inside an AI model. More parameters mean more patterns learned from data, and generally stronger performance on tasks like writing or complex reasoning. GPT-4 is widely estimated at around 1.8 trillion parameters. Ten trillion is a different order of magnitude entirely.

For comparison, Moonshot's Kimi K3, currently the largest AI model released by any Chinese company, sits at roughly 3.3 trillion parameters. We covered Kimi K3's market impact back in July in our story on how it rattled US investors. ByteDance's project would be three times that size.

What stage is this at?

The model is in pre-training: the first and longest phase of building an AI, where the system reads vast amounts of text and data to learn basic language and reasoning. Think of it as school before any specialist training begins. This phase typically takes three to six months.

After pre-training, engineers run fine-tuning, training the model on specific tasks before any public release. The 10 trillion figure isn't final either. One source said the exact size gets decided later in the process, so the headline number could still shift.

Why does this matter to ordinary people?

Larger models tend to produce better results on hard problems: complex legal documents, medical questions, multi-step coding. If ByteDance ships at this scale, it gets a serious tool to compete with US labs like Anthropic and OpenAI globally.

It also shows how fast Chinese companies are moving. On 3 August we reported that Alibaba's Qwen3 Max, a 2.4-trillion-parameter model, already rivals Anthropic's flagship on most benchmarks. ByteDance would be aiming four times higher than that.

For consumers, more competition at the frontier historically pushes prices down and quality up. ByteDance already operates the Doubao chatbot, which has tens of millions of users inside China.

Training can fail, results can disappoint, and plenty of models get shelved before anyone outside the lab sees them. But the sheer scale of the attempt tells you where ByteDance's ambitions are pointed. The company isn't chasing the middle of the pack; it's swinging at the very top.