10 trillion parameters.

If the Financial Times report is accurate, ByteDance is pretraining an AI model whose parameter scale may reach that number — more than 10× larger than all currently known public models.

The news broke in mid-August. Citing three insiders, the FT reported that ByteDance has launched super-model pretraining under a newly established tier-1 department, targeting a parameter scale that could reach up to 10 trillion. ByteDance hasn't officially confirmed this, but the organizational restructuring节奏 lends credibility to the report.

What 10 Trillion Parameters Means: A Dual Bet on Compute and Capital

To understand what 10 trillion parameters entails, we need to establish the current benchmark.

As of August 2026, the largest publicly known model is GPT-4, estimated at 1.8 to 2 trillion parameters. China's DeepSeek-V3 sits at roughly 670 billion; NVIDIA-Hopper joint efforts reached about 1.5 trillion. If ByteDance's 10 trillion figure is accurate, it would significantly outpace all known models.

But parameters are only one side of the coin. The bigger questions are: how much compute does a 10T-parameter model require? How much money? How much time?

Industry estimates suggest training a trillion-parameter model requires tens of thousands of high-end GPUs running for months to a year. At ~$30,000 per H100 GPU, ten thousand cards alone exceed $300 million — not counting power, cooling, and data center operating costs. A 10 trillion-parameter model would cost roughly 10× that.

Does ByteDance have the financial muscle? Based on available information, likely yes. ByteDance's 2025 revenue exceeded $150 billion, with profit scales sufficient to support billions in AI investment. More importantly, ByteDance already has mature compute infrastructure — from self-developed chips to large-scale data centers to a globally distributed cloud network.

The Competitive Logic of Supermodels: Is Bigger Always Better?

But does "bigger" actually mean "better"? This has been a long-debated命题 in the AI industry.

The "bigger is better" camp argues: there's a clear positive correlation between parameter scale and model capability. Larger models handle more complex tasks, generate higher-quality outputs, and process more context. From a Scaling Law perspective, as long as data and compute keep up, increasing parameters nearly always improves performance.

The opposing camp points out: the marginal returns of parameter scaling are diminishing. Once models reach a certain scale, the cost of adding more parameters far exceeds the performance gains. Additionally, super-large models face high inference costs, deployment difficulties, and staggering energy consumption — practical constraints driving the industry toward more efficient technical paths.

ByteDance's decision to announce the 10 trillion plan at this moment may reflect a deeper strategic consideration: before the final supermodel格局 crystallizes, use parameter scale to claim mindshare. The global supermodel market remains in a "land grab" phase — whoever announces the "largest model" first occupies the "most powerful model" position in users' minds.

ByteDance is simultaneously reshuffling leadership: former AI data-and-security lead Fu Yue is departing, replaced by Wang Yinglei, formerly head of the TikTok Platform Responsibility team. This personnel move sends a clear signal: ByteDance is elevating AI to a higher priority, re-integrating the entire AI chain from data security and content moderation to model training.

Whether this 10 trillion-parameter model ever trains successfully or not, ByteDance's move once again proves: the supermodel race has never stopped. Even after DeepSeek's low-price revolution shook the market and OpenAI and Anthropic pulled ahead in profitability, Chinese tech giants are still pushing toward higher numbers.

See you tomorrow.