ByteDance Plans Development of Model Exceeding 50 Trillion Parameters

ByteDance is in discussions to develop a large-scale model with training parameters exceeding 50 trillion, which would surpass the current largest models in China, Alibaba's Qwen 3.8-Max and Moonlight K3, with 24 trillion and 28 trillion parameters respectively. This initiative is still in its early stages and does not guarantee a final product.

The project will be led by Xiang Liang, head of the Seed Foundation, in collaboration with Shen Ke, who oversees pre-training data for large language models. The Seed team is currently reorganizing to better allocate resources and responsibilities for this ambitious project.

In a recent company-wide meeting, ByteDance founder Zhang Yiming emphasized the challenges of training large models and reassured the team that it is acceptable to take time to achieve their goals. He highlighted the importance of integrating various resources, including Volcano Engine and Feishu, to enhance computational power and data capabilities. Zhang also expressed a commitment to increasing investment in AI while advocating for a fundamental approach to overcoming barriers in artificial general intelligence (AGI).

Despite recent successes in multimodal performance with products like Seedance 2.0 and Seedream, the market response to the Seed 2.0 language model has been underwhelming, particularly due to its coding capabilities. In response, ByteDance has hired Guo Daye to focus on coding and has consolidated resources to address these challenges. The company aims to achieve a competitive edge by developing a larger model while fostering collaboration across departments.

© 2026 KLEA News. All Rights Reserved. This article is provided for informational purposes only. It is not offered or intended to be used as legal, tax, investment, financial, or other advice.

Source: KLEA News