10 月 22 日，科大讯飞披露最新开发者生态数据。目前讯飞开放平台开发者总数已经达到 1187 万，2026 年大模型相关企业开发者同比增长超过 104%。此外，按照讯飞以单次调用互动数超过 1 万衡量复杂任务的口径，平台上使用大模型处理复杂任务的用户占比，已经从此前不足 10%快速提升至 60%。
这意味着，现在企业开始把长链路、多任务和复杂上下文工作交给 Agent，希望它能够真正进入业务流程并持续执行。
围绕这一变化，讯飞此次一口气推出和升级了几套彼此衔接的体系：星辰 TaaS 负责把底层模型和 Token 组织成可交付的行业方案；Agent OS 负责不同员工、智能体和业务系统之间的上下文共享与任务调度；APA 进一步深入 CRM、ERP、OA 等企业软件，把传统依赖人工编排的流程转化为 Agent 可以直接调用的技能；XFYUN CLI 则开始把开放平台本身改造成更适合 Agent 调用的开发基础设施。
星辰 TaaS：不再只卖 Token，而是直接交付 Solution
“让 Token 不空转。”讯飞反复强调。
过去企业使用大模型，往往停留在购买 Token 或者调用 API 阶段。但对于真正的行业客户来说，模型调用本身并不是价值，只有当一次 Token 消耗最终完成一项业务任务，这些调用才真正变成生产力。
星辰 TaaS 可能是讯飞最明确体现这一商业化思路产品之一。其全称是 Token as a Service/Solution 体系，核心目标是把模型调用最终转换成可以交付的业务结果。
其底层由星辰 MaaS 平台提供星火和其他开源模型；中间层通过模型调度、智能体引擎、知识工程、工具编排和运行治理，把模型能力组织起来；应用层再进一步形成同传翻译、数字员工、办公提效、知识问答和智能营销等应用，并最终进入企业服务、金融、医疗、教育和能源等行业。
Agent OS：让不同智能体“接力干活”
企业中的复杂任务通常不会只发生在一个系统里。一项业务可能同时涉及 CRM、ERP、OA、合同系统、项目管理系统以及不同部门员工。即便每一个环节都部署一个智能体，如果上下文无法在不同 Agent 和不同系统之间流动，仍然很难真正形成自动化闭环。
此次讯飞开放平台首次推出 Agent OS，并联合 APA 业务流程自动化能力构建企业级 Agent 体系。其中，Agent OS 负责组织协同和任务调度，通过“场景包”把员工、智能体和业务系统组织起来，同时通过组织级上下文工程和持续进化机制，让技能和流程不断优化。
按照讯飞公布的实践数据，引入 Agent OS 后，跨系统协同效率可以提升 90%，75%以上流程能够实现自主流转。
APA 则承担具体执行。传统 RPA 需要技术人员手动编排“点击哪里、读取什么、再填写什么”。讯飞此次展示的 APA，则可以通过录屏操作或者自然语言描述，让 AI 直接理解企业软件，把填写表单、查询库存、发起审批等操作自动封装成技能，再由 Agent 按需调用。
这种多 Agent 协同已经开始进入真实业务。在国家能源集团的招采场景中，讯飞将评标任务拆分给材料归集、文档解析、标准提取和不同模块评审等多个专家智能体，再由总智能体统一协调。通过“无引用不结论、无材料不判断”等约束机制，每一个结论都要求能够追溯到原始文件，同时关键环节仍保留人类专家确认。
讯飞公布的数据显示，目前招采智能评审准确率达到 97%左右，相关能力已经服务 50 多家央国企，累计 AI 评审超过 50 万次。
在广告营销场景，讯飞则把诊断师、统筹师、策略师和巡检师等不同数字员工组成 Agent Teams，持续监控投放数据、发现问题并动态调整策略。在某电商 KA 客户海外业务实践中，引入“超级效果优化师”后，人均操盘账户投放规模提升 40%，ROAS 提升 50%。
一句命令调用 700 多项 AI 能力，开放平台也开始 Agent 化
当 Agent 成为开发者的重要“同事”，过去面向人设计的开发平台也需要重新设计。
讯飞此次推出 XFYUN CLI 开发工作台，开发者通过一句终端命令即可调用 700 多项 AI 原子能力和 200 多项智能体适配能力，把过去复杂的 API 接入、鉴权、环境配置等过程进一步隐藏到平台底层。
这意味着开发者不再需要逐项搜索文档、配置接口，而可以直接让 Agent 寻找和调用所需能力。
AIUI Code 则进一步让开发者通过自然语言描述需求，由 Agent 调用插件辅助完成开发；“交互研发数字员工”甚至可以通过语音指令直接修改机器人配置，实现安装、配置、即时更新以及不同场景的快速适配。
讯飞还升级了 GuideX 软硬一体化体系。其中，AStudio 作为 AI 生产力工作台负责接收需求，GuideX 则作为集成 Agent、虚拟人和多模态能力的交互智能平台，完成后续配置、组装和交付。同时，平台还围绕 Skill、MCP、连接器、CLI 以及行业知识建立可复用 Agent 资源，试图把过去高度定制化的 AI 项目逐步变成标准化、工程化和可规模复制的开发体系。
社区开始反过来“改模型”
除了商业化落地，讯飞也在进一步扩大开源投入。
今年 9 月，讯飞发布星火 X2.5-4B。模型开源之后，很快有社区开发者基于原版增加量化、进行代码能力校准，并修复对话模板，使模型更适合作为本地编程助手。包括这些社区版本在内，星火 X2.5-4B 一个月下载量达到 256 万次。
在其看来，过去模型厂商与开发者之间更多是一种单向关系：厂商发布模型，开发者调用模型。开源之后，这种关系开始变成双向迭代：模型厂商提供基础能力，开发者根据具体场景对模型进行量化、调优和工具链适配，社区积累下来的实践又可以进一步反哺官方模型和产品。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
On 22 October, Kodak iFlytek released the latest developers ' ecological data. The total number of developers of the iFlytek Open Platform has now reached 1,187,000, and the number of LLM-related business developers increased by more than 104 per cent in 2026. In addition, the percentage of users on the platform who use LLM for complex tasks measured by the single number of iFlytek calls over 10,000 times has risen rapidly from less than 10 per cent to 60 per cent.
This means that the enterprise is now beginning to hand over long-link, multi-task and complex context to Agent, hoping that it can truly enter the business process and sustain it.
In response to this change, iFlytek introduced and upgraded several interconnected systems: Star TaaS, which organizes the bottom model and Token into a deliverable industry programme; Agent OS, which is responsible for the sharing and tasking of context between different employees, AI agent and the business system; APA, which goes further into enterprise software such as CRM, ERP, OA, which converts traditional processes relying on manual layout into skills that Agent can call directly; and XFYUN CLI, which begins to transform open Z89Z into a more appropriate development infrastructure for Agent.
Star TaaS: No more selling Token, but delivering Solution
"Let Token not turn empty." iFlytek repeatedly stressed.
In the past, enterprises used LLM, often staying at the stage of purchasing Token or calling API. For real industry clients, however, model calls are not per se value, and they become real productivity only when a Token consumption finally completes a business assignment.
The star Taas is probably one of the products of iFlytek that most clearly reflects this commercial thinking. Its full name is Token as a Service/Solution system, with the core objective of eventually converting the model into deliverables.
The base is provided by the Star Maas platform, which provides fire and other open-source models; the intermediate layer organizes modelling capabilities through model movement, AI agent engines, intellectual engineering, tool organization and operational governance; and the application layer further develops applications such as simultaneous translation, digital staff, office efficiency, knowledge question and answer, and smart marketing, and eventually into business services, finance, health, education and energy.
Agent OS: Get different AI agent "work"
Complex tasks in an enterprise do not usually occur in a single system. An operation may involve both CRM, ERP, OA, contract systems, project management systems and staff in different departments. Even if a AI agent were to be deployed at every link, it would be very difficult to really create an automated closed ring if context could not move between different Agent and different systems.
This time, the iFlytek Open Platform launched the first Agent OS and joined the APA business process automation capability to build the Enterprise-level Agent system. Among them, Agent OS is responsible for organizing teamwork and tasking, organizing staff, AI agent and business systems through a “situation package”, while optimizing skills and processes through organizational context engineering and continuous evolution mechanisms.
According to the practice data published by iFlytek, the introduction of Agent OS could increase inter-system synergies by 90 per cent and more than 75 per cent of processes could be mainstreamed.
APA undertakes specific implementation. Traditional RPA requires technicians to organize manually “where to click, what to read, what to fill”. The APA presented by iFlytek, which allows AI to directly understand the enterprise software by means of a video operation or a natural language description, automatically encapsulates the filling out of forms, searching for inventories, initiating approvals, etc., into skills and is then called by Agent on demand.
This multi-Agent synergy has started into real business. In the National Energy Group ' s collection scene, iFlytek gave the bid evaluation task split to various experts, such as AI agent, for material sorting, document analysis, standard extraction and review of different modules, coordinated by the total AI agent. Through binding mechanisms such as “no reference, no material, no judgement”, each conclusion requires that it can be traced back to the original document, while the key elements remain as confirmed by human experts.
According to data published by iFlytek, the correct rate of recruitment for smart evaluations is currently around 97 per cent, the relevant capabilities have been served by more than 50 Qong State enterprises, and the cumulative number of AI reviews exceeds 500,000.
In the advertising arena, iFlytek makes up different numbers of employees, such as diagnostics, conciliators, strategists and inspectors, who continuously monitor the placement of data, detect problems and evolve strategies. With the introduction of the Super Effects Optimist in the business practices of an electrician KA client overseas, the per capita impact account was increased by 40 per cent and ROAS by 50 per cent.
One-word command calls over 700 AI capabilities and open platforms start Agentization
When Agent became an important “colleague” of developers, the development platform that had been designed for people needed to be redesigned.
This time, iFlytek has launched the XFYUN CLI development desk, which allows developers to call more than 700 AI atomic capabilities and 200 AI agent adaptation capabilities by a terminal command, further concealing the complex API access, assurance, environmental configuration processes of the past to the bottom of the platform.
This means that the developers no longer need to search for documents and configure interfaces on a case-by-case basis, but can direct Agent to find and call the required capabilities.
APUI Code further allows developers to describe needs in natural languages, supported by the Agent Call Plugin; "interactive development of digital staff" can even modify robotic configurations directly through voice commands to achieve installation, configuration, instant updates and rapid adaptation of different scenarios.
iFlytek also upgraded the GuideX soft and hard integration system. Of these, AStudio is responsible for receiving demand as the AI Productivity Workstation, while GuideX serves as an interactive smart platform for integrating Agent, virtuals and multimodular capabilities to complete subsequent configuration, assembly and delivery. At the same time, the Platform is also building reusable Agent resources around Skill, MCP, connectors, CLI and industry knowledge in an attempt to gradually transform past highly customized AI projects into standardized, engineering and replicable development systems.
Communities begin to “modify” in turn
In addition to commercial landings, iFlytek is further expanding open source inputs.
In September this year, iFlytekpublish fire X2.5-4B. Following the opening of the model, there will soon be community developers who add quantization to the original version, calibrate code capabilities and repair the dialogue template to make the model more suitable for local programming assistants. Including these community versions, the fire X2.5-4B was downloaded 2.56 million times a month.
In its view, in the past there was more a one-way relationship between model manufacturers and developers: the manufacturer publish model, the developer called the model. After the open source, the relationship began to be a two-way iterative one: model manufacturers provided the basic capacity, developers adapted the model to the specific scenes, adapted the quantization, adapted the tool chain, and community practices built up to feed official models and products further.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.