Introduction

According to reliable reports published by the Financial Times last Friday, citing informed sources and technical industry experts, Chinese technology giant ByteDance is currently working in secret to train a massive and highly advanced artificial intelligence model containing up to 10 trillion code parameters. This enormous training scale represents a bold and unprecedented step that could soon bring this new model very close in its generative capabilities and analytical efficiency to the advanced and complex Mythos system developed by leading American company Anthropic. This ambitious technological project is currently in the intensive pre-training phase, a complex and critical stage expected to take between three to six continuous months of intensive computational processing across a giant network of servers. If the Chinese company succeeds in completing this project, the size of its model will be more than three times the size of the Kimi k3 model developed by rival Chinese startup Moonshot AI, which contains only 2.8 trillion parameters. Upon completion at this massive and domestically unprecedented scale, this model will undoubtedly be classified as one of the largest and most complex artificial intelligence systems developed in human history to date.

• Massive scale and competitive context of language models
• Reliance on independent development and management directives
• The technical arms race among major global powers
• Frequently asked questions

Massive scale and competitive context of language models

Although American company Anthropic does not typically or publicly disclose the parameter counts of its advanced models to protect its sensitive trade secrets, estimates by industry experts and analysts indicate that its globally leading Mythos 5 system contains about 8 trillion parameters, while its other system, Fable 5, contains about 5 trillion parameters. Consequently, ByteDance’s large and announced ambitions will place it in an advanced position, perhaps at the forefront of current global artificial intelligence technological development. Parameters programmatically represent precise numerical values and neural connections that the model learns and continuously adjusts during the intensive training process of the neural network, enabling it to recognize highly complex linguistic patterns and generate accurate, logical, and detailed responses that mimic human understanding. However, specialized researchers always and firmly emphasize that the total size or numerical scale of parameters alone does not absolutely determine the final capability of the model, as other crucial technical factors play a pivotal role, such as the high quality and superior purity of training data, the architectural design of neural networks, and continuous algorithmic optimization techniques. Before the emergence of the Kimi k3 model, the largest models in China included domestic models such as LongCat 2.0 by Meituan and V4 Pro by DeepSeek, each possessing only about 1.6 trillion parameters, clearly highlighting the magnitude of the leap ByteDance is now attempting to achieve.

Reliance on independent development and management directives

This important news report published by the British newspaper coincides with very strict and decisive internal directives issued by ByteDance founder Zhang Yiming, who asked all employees and research teams at the company to immediately and absolutely stop trying to improve and develop their models by distilling or cloning the outputs of competing Western systems in the market. Instead, the founder ordered a preference for building completely independent and authentic long-term research and development capabilities, entirely steering away from short-term technological gains reliant on the efforts of others, reflecting a genuine desire to achieve tangible technological sovereignty. These strict administrative orders follow strong and direct accusations leveled by Anthropic against several prominent Chinese artificial intelligence companies for widely using the outputs of its Claude language model to secretly train their competing models. In the context of its own intensive efforts, ByteDance established a special and advanced research team in early 2023, which quickly expanded to currently employ about 2,000 experts, researchers, and engineers distributed across company offices throughout China and abroad. This team’s capabilities were significantly boosted recently by the joining of prominent technology expert Wu Yonghui, who previously worked for long periods at American companies Google and DeepMind, to head this research team in February 2025 in order to guide and direct these independent efforts with global expertise toward broader horizons.

The technical arms race among major global powers

This accelerated and major development clearly highlights the intensification of fierce and open technological competition between Chinese and American technology companies for complete control of the vanguard of the advanced global artificial intelligence sector. In the past few weeks, some recent Chinese models developed by rising companies such as Moonshot and the Alibaba Group received extremely positive reviews in independent comparative evaluations, trailing only by a very narrow margin behind the Fable 5 model developed by Anthropic in certain technical areas related to logic and complex reasoning. Unlike many other emerging Chinese companies that favor and support open-source artificial intelligence technologies to share knowledge and attract the global developer community, most of ByteDance’s primary and advanced models remain closed-source and restricted to internal commercial use. Its advanced chatbot known as Doubao is currently the most popular and widely used artificial intelligence model across all of China, boasting a massive user base exceeding 324 million active users interacting with it monthly. Furthermore, its video generation and creation model named SeedDance is among the most innovative, advanced, and progressive models globally in terms of accuracy and realism. As of the time of writing this detailed report, ByteDance has not responded to requests for comment and clarification submitted by the Financial Times regarding the scale of its new secret project and its launch date.

Frequently asked questions

Question: What is the size of the new language model currently being developed by ByteDance?

Answer: The company is working hard to train a giant and highly advanced artificial intelligence model containing 10 trillion code parameters, making it one of the largest models under development in the world and surpassing its domestic competitors by multiple folds in capacity.

Question: How does this massive Chinese model compare to models from prominent American companies in the field?

Answer: If the project is successfully completed, the size of the new model will be very close to or even numerically superior to the largest systems of leading American company Anthropic, whose strongest system Mythos 5 is estimated to have about 8 trillion parameters.

Question: What are the new administrative directives imposed by the company founder on working research teams?

Answer: The company founder issued strict orders to halt the use of competing foreign model outputs to improve their systems, relying instead wholly and exclusively on the development of independent, long-term research and capabilities to avoid accusations of technical cloning.

Question: Does the massive size of the model and the increase in parameter count guarantee its actual superiority in overall AI performance?

Answer: No, specialized researchers emphasize that increasing the number of parameters alone is never sufficient; it must always be coupled with high-quality and heavily filtered training data, innovative neural network software architecture, and continuous algorithmic optimization processes to ensure accurate and effective results.