ByteDance has postponed the release of its Doubao 2.2 large model, which had been scheduled for August, as it extends pre-training and post-training work to strengthen coding, tool invocation and agent capabilities. Sources close to the company said the goal is to move the Seed model into the industry’s first tier for coding before year-end, benchmarking against Zhipu GLM-5.2 and Kimi K3. The company has recently been rolling out frequent small-step coding iterations, accelerating team adjustments and recruiting external talent. On August 20, ByteDance reorganized its Seed foundation model unit into four first-level departments: Pretrain Data, Horizon RL, Product Posttrain-Work, and Product Posttrain-Chat, aiming to improve long-context handling, task planning and long-duration task execution.