What China’s L3 AI Phone Rating Really Means

China’s new AI-terminal grading system is best understood as a capability benchmark for complete devices, not as a “self-driving” scale for phones. The first results, released on July 17, 2026, listed 11 mobile terminals at L3: nine smartphones and two tablets. They show that particular submitted configurations can provide bounded, user-supervised assistance—not that the devices can act without oversight across every app or service. 25

What GB/Z 177—2026 measures

The framework is called GB/Z 177—2026, Intelligence Grading of Artificial Intelligence Terminals. The “GB/Z” designation identifies a national standardization guidance document, so the system provides a common way to evaluate AI capability rather than a compulsory condition for selling a phone. 156

It assesses the terminal as a system. That includes the model, operating system, app and tool interfaces, memory, permissions, and the relationship between on-device and cloud processing—not simply the size or conversational quality of an underlying language model. 519

The framework organizes capability around five dimensions:

• Perception: understanding signals such as text, voice, images and other inputs.
• Cognition: interpreting intent, reasoning and planning.
• Execution: selecting tools and completing actions.
• Memory: retaining relevant short- and long-term context.
• Learning: adapting to changing situations and user preferences. 519

Those dimensions are further divided into 14 sub-capabilities, including user perception, task planning, tool invocation, short- and long-term memory, and contextual adaptation. 19

The progression from L1 to L4

The four labels describe an increasing ability to move from reacting to instructions toward coordinating work across systems. The first three levels have defined capability requirements; L4 is named in the framework but remains to be fully specified. 52125

Level	What it means in practice
L1 — Response	Understands a clear instruction and performs a straightforward, usually single-step action, such as opening a function or setting a reminder.
L2 — Tool	Understands a simpler intent, performs limited reasoning and follows predefined tools or workflows for constrained multi-step tasks. It is closer to a capable chatbot or in-app assistant than an open-ended agent. 519
L3 — Assistance	Understands a broader goal, asks for missing information, breaks work into steps, selects and orchestrates tools, handles multimodal interaction, and uses short- and long-term memory. 523
L4 — Collaboration	Intended to coordinate among agents, devices and services, but its detailed requirements and test methods are still being developed. 57

The important distinction is not that an L3 phone “thinks like a person.” It is that the terminal can connect understanding, planning, tool use, memory and action into a more complete task workflow. 519

What the first 11-device batch demonstrates

The July results included nine smartphones associated with Huawei, Motorola, Honor, vivo, OPPO, Xiaomi and Stepfun, along with two tablets. 235

The narrow conclusion is that the submitted device configurations met the current L3 test requirements. The result does not mean:

• every phone made by a listed brand is L3;
• every software version on a qualifying model has identical capabilities;
• every advertised AI-agent function works across all apps, regions or services; or
• phones absent from the first list necessarily failed.

The first batch came from vendor-submitted products rather than a random survey of the market. An L3 result is therefore a product-and-configuration rating, not a brand-wide intelligence ranking. 56

Why L3 is not a mandatory phone certification

Calling the result a “national L3 certification” can obscure what the label does and does not do. Because GB/Z 177—2026 is a guiding technical document, an L3 result is not a regulatory authorization required before a phone can be sold. It is an industry benchmark for capability and evaluation. 16

The label also should not be read as proof of universal autonomy. L3 evaluates whether a terminal can provide complex assistance under defined conditions; it does not grant the device unrestricted access to personal data, accounts, payments or every third-party app. The user remains part of the control loop, especially for consequential actions. 519

L2 chatbot versus L3 agent: a practical example

An L2 assistant might answer a question about the weather in Hangzhou, search for information or generate a suggested itinerary through a preset workflow. The conversation and task usually remain within a limited tool path or current session. 519

An L3 phone is designed to handle the broader goal behind a request. For example, after being asked to plan a weekend trip, it could:

1. ask for dates, budget and other missing constraints;
2. combine transport, weather, hotel and calendar information;
3. decide which tools are needed and sequence them;
4. remember preferences such as avoiding early flights or staying near a particular area; and
5. prepare an itinerary, calendar entry or booking while asking the user to approve sensitive disclosures, payments, reservations or other irreversible actions. 51923

That final step defines the boundary. L3 is supervised delegation, not unrestricted authority. The agent may prepare an action, but final control should return to the user when money, sensitive data or an irreversible commitment is involved. 5

Why L4 remains undefined

L4 is not simply “a more powerful L3.” Its defining challenge is coordination among independent agents, devices and services. A workable L4 test would need clear rules for:

• delegating tasks across devices and agents;
• transferring permissions or payment authority;
• verifying identity and consent;
• maintaining audit trails and security boundaries; and
• assigning responsibility when a connected action fails.

Those questions involve the phone maker, operating-system provider, app, service provider, model provider and user. Until the industry can express those relationships as repeatable and safe test cases, a detailed L4 certification method is difficult to establish. The next meaningful step may therefore depend as much on interoperability, permissions and accountability as on faster chips or larger models. 57

L3 is not the same as automotive autonomy L3

The shared L1–L4 labels are easy to misread. Automotive autonomy grades describe responsibility for the dynamic driving task, including who monitors the vehicle and when a human must take over. The AI-terminal framework instead grades capabilities such as perception, intent understanding, reasoning, tool use, memory and learning. 5

An L3 phone is therefore not equivalent to a conditionally autonomous car. Likewise, a future L4 phone would not automatically mean a device that can safely make every consequential decision without human confirmation. 5

What consumers should look for

The Huawei Pura X Max and Lenovo’s Motorola devices are useful examples of products associated with the first L3 results, but the label should not be the only buying criterion. 25

A phone’s practical agent capabilities can depend on operating-system updates, assistant and model availability, app integrations, permissions and cloud connectivity. That means useful L3-style functions do not automatically require a new hardware purchase, although newer hardware may improve factors such as speed, privacy or battery performance. 5

Before treating an L3 badge as a buying decision, check:

• which apps and services the agent can actually operate;
• what personal data and permissions it receives;
• which actions require explicit confirmation;
• whether stored memories can be reviewed or deleted; and
• whether the feature works in your language, region and accounts.

The clearest takeaway from the July batch is that AI phones are becoming measurable as task-performing systems. But L3 is a supervised assistance benchmark for specific configurations—not a blanket promise of autonomous behavior, not a mandatory market-entry approval, and not the final level of terminal intelligence. 157