Every phone maker, cloud vendor, and PC builder in China is now shipping something it calls an "AI operating system" (人工智能操作系统) — but until recently no two agreed on what that even meant. On September 24, 2026, the State Administration for Market Regulation and the Standardization Administration (国家市场监督管理总局、国家标准化管理委员会) published two national standards that try to fix that: General Technical Requirements for AI Operating Systems (《人工智能操作系统通用技术要求》, GB/Z 299-2026) and Permission Management Requirements for AI Operating Systems (《人工智能操作系统权限管理要求》, GB/Z 300-2026). Both took effect the same day and are currently in force (现行).
A launch ceremony for the two standards was held on October 13, 2026, at the New Industries Standardization Leadership Forum in Beijing, with the Ministry of Industry and Information Technology's (工业和信息化部) IT Development Department and the China Electronics Standardization Institute (中国电子技术标准化研究院) in attendance.
Key takeaways
- The two documents are GB/Z — national guidance technical documents, not mandatory GB/T standards — so they steer industry rather than legally compel it.
- GB/Z 299-2026 sets a technical reference framework across six dimensions: heterogeneous-compute scheduling, agent runtime, AI capability framework, security and privacy, multi-device collaboration, and proactive intelligence.
- GB/Z 300-2026 builds a unified permission system spanning users, apps, agents (智能体), skills (技能), and tools (工具) — essentially a safety gate for who can do what inside an AI OS.
- They were drafted by the China Electronics Standardization Institute with 60-plus organizations including Huawei, ZTE, Xiaomi, Tongxin (统信), Kylin (麒麟), China Telecom's Tianyi Cloud (天翼云), OPPO, Honor (荣耀), vivo (维沃), and Nanjing University.
What GB/Z 299 actually specifies
The general-requirements standard defines a framework for AI operating systems aimed at smart terminals and personal-use scenarios. It covers the kernel, AI runtime libraries, AI services, the system agent (系统智能体), an agent development kit, and AI security. The six technical dimensions are concrete: it asks systems to unify scheduling of heterogeneous computing (异构算力统一调度), provide an agent runtime environment (智能体运行时环境), expose an AI capability framework, embed a security-and-privacy framework, support multi-device collaboration, and enable proactive intelligence (主动智能). The scope is design, development, and testing — a reference coordinate more than a checklist.
What GB/Z 300 adds: the permission gate
If 299 answers "how do you build it," 300 answers "how do you govern it." It sets security principles and technical specs for permission management, establishing a unified system for multiple actor types — users, applications, agents, skills, and tools. It specifies identity authentication and management, permission classification and grading (权限分类与分级), permission models and mapping, full-lifecycle permission management, and scenario-based permission control. In an era where agents can call tools and chain actions across apps, that unified gate is meant to stop over-privileged agents and data misuse at the source.
Why "guidance" and not "mandatory"
The GB/Z label matters. China's national standards come in mandatory (GB), recommended (GB/T), and guidance (GB/Z) forms. Guidance documents (指导性技术文件) are typically used where technology is still moving fast and the committee wants to codify best practice without freezing innovation or imposing legal compliance costs. For a category as young as the AI OS, that is a deliberate choice: set the common language now, tighten later if the market converges. Industry players we cite did not announce certification obligations, so adoption for now is reputational and procurement-driven rather than legally mandated.
The market gap these standards fill
Until now, "AI OS" has been a marketing label more than an engineering category. Vendors bolted model runtimes, assistant layers, and device-control stacks onto existing Linux or Android bases without a shared vocabulary for what a system agent should do, how permissions should cascade to sub-agents, or how heterogeneous NPUs and GPUs should be scheduled together. That fragmentation raised integration cost for app developers and made security review ad hoc. By naming the six dimensions and the multi-actor permission model, the standards give buyers a checklist and give smaller OS vendors a target, which is precisely why a guidance document — rather than a mandate — can move the market.
Honest limitations
We confirmed the standard numbers, publication and implementation dates, the GB/Z status, the issuing bodies, and the six technical dimensions against the national standards platform (std.samr.gov.cn) and the TC28 committee site (nits.org.cn), cross-checked with provincial media. The full detailed technical clauses were not independently re-derived line-by-line; we report the committee's published summaries. Because these are guidance documents, their real impact depends on voluntary adoption and any future procurement or certification rules, which we could not verify. We did not assess compatibility with international OS or AI standards.
What readers can do now
- If you build AI OS or agent platforms in China, align designs to the 299 framework and the 300 permission model — they are becoming the de-facto procurement and interoperability baseline.
- If you buy AI devices or OSes, ask vendors whether their product maps to GB/Z 299/300; it is a fast filter for maturity.
- If you track standards, watch for a possible upgrade from GB/Z (guidance) to GB/T (recommended) once the market settles — that would shift these from steering to expectation.