字节的 Agent 产品来了。 4 月 18 日晚间，字节跳动扣子空间开启内测，定位通用 Agent。与其他类似产品如 manus 一样，扣子空间采用了邀请码制。 平台上，用户可以选择精通各项技能的「通用实习生」，也可以选择行业的「领域专家」，通过与 AI 的互动完成工作任务。 就在前一天，火山引擎刚刚面向企业市场推出 OS Agent 解决方案及 AI 云原生推理套件，要帮企业更快、更省地构建和部署 Agent 应用。 扣子有探索和规划两种模式，如果想让它一步到位输出，可以选择探索模式；如果想亲自把控每个步骤，可以选择规划模式。 虽然采用了邀请码制，但扣子显然不是走饥饿营销的路线。用户激活获得的邀请码后，创建并完成一个新任务即可获得 5 个邀请码，邀请码激 活后，还可获得更多邀请资格，多创建多获得多邀请。 上线的第一时间，极客公园就上手实测了扣子空间的 Agent 功能。可以看到，在执行的三个任务中，制定旅游攻略和一周穿搭的任务顺利完 成，但另一个专家助手的任务，却出现了 Python 脚本调用失败、 API 权限异常等 Bug，一个晚上都没能执行成功。 字节迈出了走向 Agent 时代的第一步，但距离完美 ...

Core Insights - The Volcano Engine has officially launched the Doubao 1.5 Deep Thinking Model, which utilizes the MoE architecture with a total parameter count of 200 billion and an active parameter count of 20 billion, achieving top-tier performance in multiple benchmark tests [1][3][8] Model Capabilities - Doubao 1.5 features practical capabilities such as "thinking while searching" and "visual understanding," available for enterprise users on the Volcano Ark platform [3][4] - The model can achieve a low latency of 20 milliseconds in high-concurrency scenarios, allowing it to perform searches and reasoning simultaneously [4][6] - It demonstrates visual understanding by analyzing text and image information, providing tailored recommendations based on user preferences [6][20] Performance Metrics - In various authoritative benchmark tests, Doubao 1.5's scores are comparable to OpenAI's models, particularly in mathematical tests like AIME 2024 and AIME 2025, while showing significant advantages in the ARC-AGI test [8][10] - The model scored 77.3 in the GPQA Diamond reasoning challenge, closely trailing OpenAI's models, and has shown strong performance in programming benchmarks [10] Market Position - As of March 2025, Doubao's daily token usage has exceeded 12.7 trillion, marking a threefold increase from December 2024 and a 106-fold increase from its initial launch [3] - Volcano Engine holds a 46.4% market share in China's public cloud model usage, positioning it as the market leader [3] Additional Model Upgrades - The upgraded Doubao Text-to-Image Model 3.0 can generate high-quality 2K images and is applicable in various fields such as marketing and design [11][15] - The new Doubao Visual Understanding Model enhances visual localization capabilities and supports semantic video search, making it suitable for commercial applications like security and home care [17][20] Industry Context - The competition among domestic reasoning models is intensifying, with Doubao 1.5's advancements in reasoning costs and visual understanding potentially setting the stage for the next wave of upgrades in the industry [21]

Core Insights - Volcano Engine has launched Doubao 1.5 deep thinking model for the enterprise market, along with upgrades to the text-to-image model 3.0 and visual understanding model, as well as the OS Agent solution and AI cloud-native inference suite to help enterprises build and deploy agent applications more efficiently [1] Summary by Categories - **Product Launches** - Doubao 1.5 deep thinking model has been introduced for enterprise applications [1] - Upgrades to the text-to-image model 3.0 and visual understanding model have been announced [1] - OS Agent solution and AI cloud-native inference suite are now available to assist enterprises [1] - **Performance Metrics** - As of March 2025, the daily token usage for Doubao large model has reached 12.7 trillion, representing over a hundredfold increase since its launch in May of the previous year [1]