On September 22, the inaugural Immersive Consumption Carnival and the 2026 AI Terminal Innovation and Development Conference opened in the Beijing Economic-Technological Development Area (Beijing Yizhuang). Tao Hongzhi, Chief Standards Officer of Lenovo, attended the conference and delivered a speech. Tao pointed out that the terminal industry has entered the third generational revolution, officially stepping into the era of AI terminals.
Tao stated that the two previous generational revolutions in the terminal industry gave birth to two types of terminals. The first is the functional terminal, where terminal products can only perform local computing and standalone general-purpose computing through the local OS. The second is the internet terminal. After 2000, internet terminals gradually emerged, enabling online services, but still failing to reach the level of intelligent terminals. Since 2020, the advancement of large models and AI (Artificial Intelligence) technology has brought new opportunities for the third generational revolution in the terminal industry.
Given the growth and widespread adoption of AI terminals, standardization efforts are also deepening. In May 2026, the national standard series "Intelligence Grading of AI Terminals" was released, classifying AI terminals into L1 Responsive, L2 Tool, L3 Assistant, and L4 Collaborative levels based on their intelligence capabilities. Among these, L3 is currently the highest testing rating.
Tao noted that multiple Lenovo computers were selected for the first batch of L3-level terminal product catalog. He further elaborated on the evaluation principles of the national standards for AI PCs. This series of national standards sets requirements for five capabilities and two capability scenarios for microcomputers, namely perception, cognition, execution, memory, and learning capabilities, as well as on-device capabilities and device-cloud collaboration scenarios. During the speech, he also shared Lenovo's extensive practices in these areas.
Perception capabilities include the ability to perceive user information such as text, audio, video, faces, and voiceprints; device information such as hardware, software, and peripheral devices; and environmental information such as time, weather, and time zones. For a long time, the perceived information for PCs has mainly come from mice and displays, resulting in lower capabilities and work efficiency. Addressing this issue, Tao stated that Lenovo has launched solutions such as hardware sensor arrays. Through continuous or intermittent state perception, these solutions identify the user's context and work requirements, endowing computers with system and context perception capabilities, thereby enabling proactive context triggering and automatic task execution.
Cognition capabilities encompass understanding, reasoning, and planning. Tao mentioned that Lenovo has conducted meticulous intent classification, categorizing human intents and PC work intents into 27 major categories and 428 subcategories, assisting PCs in performing complex intent understanding tasks. In terms of reasoning capabilities, through the process design of full independent storage across device, edge, and cloud, hierarchical compression, and on-demand recall, a 10x expansion of the virtual context window has been achieved.
Memory capabilities include short-term and long-term memory. Tao shared Lenovo's hierarchical memory technology, which utilizes episodic memory, semantic memory, and procedural memory technologies to organize short-term memory into long-term memory, and even personalized memory graphs. This is then input into the Agent execution environment, thereby achieving a dual improvement in accurate memory recall and behavior planning capabilities. Tao stated that through such technologies, Token consumption can be reduced by 60% to 95%, saving 60% of the time.
Execution capabilities include tool invocation, content generation, and interconnected collaboration. Tao believes that looking towards future L4-level intelligence, AI terminals are expected to further enhance their autonomous execution and autonomous operation capabilities, taking on more automated tasks. He shared Lenovo's Tianxi Harness architecture. It is reported that Tianxi Harness performs exceptionally well in long-horizon tasks, capable of executing over 200 rounds of long-horizon tasks.
Learning capabilities are mainly reflected in context adaptation. Tao stated that the advancement of AI technology requires increasingly broader contexts, better tool execution results, and continuously iterated understanding of personalized needs. Lenovo aims to endow AI PCs with the individual self-evolution capability to automatically learn and generate exclusive Skills, as well as the group self-evolution capability to extract universal skills, thereby meeting users' essential needs.
Tao also emphasized that intelligent terminals must remain vigilant about security issues. He believes that the more powerful the capabilities of intelligent terminals, the higher the requirements for the confidentiality and security of personal privacy data such as knowledge and preferences. For L4 and higher-level intelligence, in addition to achieving capability upgrades and full-scenario autonomy across device and cloud, it is still necessary to adhere to full-chain privacy computing and autonomous permission management to build a security baseline for advanced intelligence.
Looking to the future, Tao stated that the AI integration of OS components is not the endpoint. He looks forward to reconstructing the operating system through AI-native technology to drive the continuous development of AI terminals.
Author | Chen Cun
Editor | Zhang Xinyi
Art Designer | Ma Liya Supervisor | Lian Xiaodong