ByteDance Delays Doubao Large Model 2.2 Launch to Sharpen Coding Capabilities

On August 30, ByteDance announced that the release of its Doubao large model version 2.2, originally scheduled for August, has been postponed. According to sources close to ByteDance, the delay stems from the company’s plan to enhance the model’s capabilities in programming, tool calling, and agent functionalities through more extensive pre-training and post-training. This decision will extend the development cycle to achieve a more significant performance leap. Currently, ByteDance’s Seed team is simultaneously working on two fronts: advancing the development of its next-generation ultra-large model while maintaining regular iterations of the existing model. Doubao 2.2 is positioned as a critical interim upgrade between generations, aiming to specifically address coding-related challenges without waiting for the completion of the next-generation model—and avoiding mere incremental updates. At last month’s all-hands meeting of the Seed team, ByteDance founder Zhang Yiming stated that even if the company temporarily lags behind competitors, it will not adopt AI distillation techniques to improve its models. CEO Liang Rubo emphasized at the company’s mid-year all-staff meeting for 2026 that ByteDance remains committed to independently developing its large language models, strengthening foundational capabilities, accepting short-term setbacks, and focusing on long-term optimization. The company plans to bring the Seed model’s coding capabilities up to the industry-impact level of models like Zhipu’s GLM-5.2 and Kimi-K3 by the end of 2026, listing this as one of its key objectives for the year.