所见，即行动Vision into Action.

Mini 与 Pro 延续 Nex-N2 的多模态底座，重点强化计算机使用、网页浏览与视觉辅助的能动性，让视觉理解与环境交互共同服务于真实任务。Max 基于 DeepSeek-V4-Pro-Base，采用 1.6T 参数的 MoE 纯文本架构。我们首次在万亿参数规模上开展完整后训练，探索更大模型在复杂 Agent 任务中的能力。Mini and Pro build on Nex-N2’s multimodal base, improving computer use, web browsing, and visually assisted agency to connect visual understanding with real-world action. Max builds on DeepSeek-V4-Pro-Base, a 1.6-trillion-parameter, text-only MoE base. Our first full post-training effort at this scale explores larger models’ capabilities on complex agent tasks.

面向真实环境中的长程任务，Nex-N2.5 依靠视觉反馈持续行动与自我修正，操作计算机、浏览器并运行和测试程序。视觉成为 Agent 感知环境、验证结果与推进任务的重要接口，让模型能够依据环境变化调整后续行动。For real-world, long-horizon tasks, Nex-N2.5 uses visual feedback to act and self-correct, operate computers and browsers, and run and test code. Vision serves as an interface for perceiving the environment, verifying results, and advancing tasks, helping agents adapt their next actions to changes they observe.

我们进一步扩充 Agent 训练环境、任务类型与生产力场景，结合多样化任务和环境反馈，开展系统性后训练。训练强化了模型在科学研究、知识工作和复杂生产力任务中的能力，也为更大规模模型的能动性训练积累实践经验。We expanded agent training environments, task types, and productivity scenarios, using diverse tasks and environmental feedback for systematic post-training. This strengthens scientific research, knowledge work, and complex productivity capabilities, while building experience in agency training for larger models.