Perceive Beating AI News Flash: ByteDance has postponed the release of DoughV2.2, originally planned for August. ByteDance is preparing to spend more time on training, with a focus on enhancing Coding, tool invocation, and Agent.

One of ByteDance's goals set for Seed this year is to elevate its Coding to a top-tier level. Internally, there is a hope to achieve effects similar to GLM-5.2 and Kimi-K3 by the end of the year, allowing developers to truly recognize ByteDance's Coding model.

In August, Seed just completed a major restructuring. The teams originally divided into Text, Voice, Code, and Vision have been reorganized into four new departments: Pretrain Data, Horizon RL, Product Posttrain-Work, and Product Posttrain-Chat. Horizon RL will be responsible for post-training Coding, while the Work team will focus on tool invocation, GUI operations, and long task execution.

ByteDance has recently been actively hiring in this area. The official website is recruiting Code Agents, General Agents, and Reinforcement Learning Algorithm Engineers. A Multi-Agent Harness position even explicitly states the requirement to research Coding Agents such as Claude Code and Codex, and to build a continuous-running multi-agent and reinforcement learning environment.

When Seed 2.1 was released in June, it already emphasized Coding, but ByteDance apparently felt it was not sufficient. This time, DoughV2.2 would rather be delayed to ensure that Coding is significantly improved before release.