Doubao 2.2 Release Postponed, ByteDance Strengthens Coding and Agent Capabilities
- ByteDance has postponed the release of its Doubao large model 2.2, originally planned for August.
- The company extended pre-training and post-training cycles to strengthen the Seed model's capabilities in coding, tool invocation, and Agent functions.
- ByteDance aims to bring its coding level to industry first-tier by year-end, benchmarking against Zhipu GLM-5.2 and Kimi K3.
- The Seed foundation model department was reorganized on August 20 into four first-level departments focusing on pretrain data, reinforcement learning, and product post-training.
- ByteDance is iterating coding features frequently and recruiting external talent to accelerate the development.
ByteDance has postponed the release of Doubao 2.2, its large language model originally scheduled for August release. According to sources close to the company, the delay is strategic, aimed at extending both pre-training and post-training cycles to enhance the Seed model's performance.
The company is prioritizing improvements in three key areas: coding capabilities, tool invocation, and Agent functionality. ByteDance has set an ambitious goal to elevate its coding performance to industry first-tier standards by the end of the year, using Zhipu GLM-5.2 and Kimi K3 as competitive benchmarks.
To support this initiative, ByteDance reorganized its Seed foundation model department on August 20, creating four new first-level departments: Pretrain Data, Horizon RL, Product Posttrain-Work, and Product Posttrain-Chat. These structural changes aim to improve long context handling, task planning, and long-duration task execution capabilities. The company is also accelerating development through frequent iterative updates to coding features and external talent recruitment.
豆包2.2发布延期，字节跳动加强编码和Agent能力
字节跳动推迟了豆包2.2的发布，该版本原定于8月发布。据接近公司的消息人士透露，这一延期是战略性的，旨在延长预训练和后训练周期，以增强Seed模型的性能。
公司正优先改进三个关键领域：编码能力、工具调用和Agent功能。字节跳动设定了雄心勃勃的目标，在年底前将其编码性能提升至业界一流水准，以智谱GLM-5.2和Kimi K3作为竞争基准。
为支持这一举措，字节跳动于8月20日重组了Seed基础模型部门，创建了四个新的一级部门：预训练数据、Horizon强化学习、产品后训练-工作和产品后训练-对话。这些结构调整旨在改进长上下文处理、任务规划和长时间任务执行能力。公司还通过频繁迭代编码功能和外部人才招聘来加快开发进度。