Lenovo reshapes how individuals and enterprises work with AI-native devices and enterprise AI agents

Lenovo Group is advancing its AI-native transformation through personal terminals and enterprise agents. Tianxi calls tools to complete multi-step tasks, the AI Host preserves local knowledge and long-term memory, and Lexiang connects product, order, and service systems; according to Lenovo China's 2025/26 fiscal year practice data, Lexiang's sales conversion rate is 23% higher than that of the traditional official website. Moonshot AI and DeepSeek, meanwhile, organize product delivery and their own R&D around core models, presenting different paths by which domestic enterprises are reconstructing their businesses with AI.

Tianxi can now write reports, create presentations, and generate web pages. After the user states a requirement, the agent retrieves materials, organizes files, and calls relevant applications to complete the work; if the user raises new requirements midway, Tianxi can adjust accordingly.

Tianxi 4.3 adds full-duplex natural voice interaction, allowing users to interrupt the conversation at any time or pause and then continue expressing themselves. Super Mode supports cross-device control and local data processing, extending voice interaction into the execution of specific tasks.

This design covers terminals including PCs, phones, and tablets. Among the office and learning products Lenovo showcased in September, the Zhaoyang X7 supports AI meeting simultaneous interpretation and AI Notes, while the Legion Y900 tablet supports handwriting input, note organization, and intelligent document parsing. Tasks such as meetings, material reading, and content creation correspond to the input methods and computing support of different devices.

The AI Host provides device support for agents to run resident. Lenovo positions it as an AI-native edge device that can preserve local knowledge and long-term memory, carry high-frequency tasks, and coordinate with computers, phones, and tablets. Knowledge and context accumulated in personal use now have a local carrier that can be called upon continuously.

At the terminal conference in September, Lenovo Chief Standards Officer Tao Hongzhi listed complex intent understanding, hierarchical bionic memory, and long-horizon task execution as technical priorities. Personal AI adapts to users' usage habits through continuous learning and the accumulation of personalized skills, and calls on existing knowledge for subsequent tasks.

Lenovo Lexiang extends from product consultation to transaction and after-sales execution

In Lenovo China's 2025/26 fiscal year practice data, Lexiang's sales conversion rate is 23% higher than that of the traditional official website. Lexiang 4.0, launched on June 30, further expands task execution capabilities across product, order, and service systems.

In Lenovo China's agent network, Lexiang serves customers and the public, Tongxiang serves partners, and Leda serves employees; the three types of entry points coordinate with more than 100 scenario agents and continuously accumulated Skills, each addressing the needs of different parties.

In Lexiang 4.0, customers can express product consultation, product recommendation, order placement, or after-sales repair needs through natural language. Lenovo has extended its original knowledge base into a Skills system capable of invoking business actions; after providing answers, the agent can also operate relevant business systems to continue matching needs and handling services.

A customer's needs may involve product solutions, smart store experiences, and membership benefits within a single interaction. Lexiang presents the corresponding content and handles executable matters on the customer's behalf; needs added by the customer during the consultation can also continue into task processing.

Supporting execution is Lenovo's self-developed Phoenix agent runtime layer. The master control agent understands requirements, breaks down tasks, and chooses serial or parallel collaboration based on task characteristics; specialized agents connect to systems such as products, orders, and services, and relay the execution of corresponding work. Requirement understanding, task scheduling, and system operation are divided among different agents.

Lexiang handles high-frequency, repetitive product questions, while human customer service focuses more on complex matters, continuing to provide customers with full-cycle service. In enterprise purchasing scenarios, the agent handles consultation responses and online selection, and connects with smart store services, based on which customer service staff adjust their division of work.

Lexiang 4.0 also introduces hybrid memory and reflection mechanisms, incorporating operational feedback, problem identification, strategy optimization, and capability accumulation into product iteration. Business teams define operational goals and rules, while IT teams handle interfaces, security, and operational support, continuously addressing problems in real services.

Lenovo's enterprise agent offerings are delivered through methods such as platform subscriptions and on-site customer development. Baiying AI 3.0, launched in May, adopts a cloud-based, subscribable platform form to provide product services for enterprises using agents; Qingtian AI supports agent development, operations, and knowledge governance.

In the AI Foundry architecture announced in April, the agent library, model platform, agent development and operations platform, and data and knowledge governance platform constitute the main modules. In the September showcase, Lenovo further presented Lenovo Baiying, Lenovo Qingtian, and full-cycle AI services with Qingtian AI as the engine, linking the development platform with customer-facing services.

The Foshan municipal government is leading the construction of a digital and intelligent service platform for manufacturing, with Lenovo providing hybrid AI technical services. The platform has entered the delivery stage, building dedicated datasets around local industries to form modular agents covering R&D, production, quality inspection, supply chain, and sales. Process parameters, quality standards, and equipment patterns from industries such as advanced manufacturing, pan-home furnishing, textiles, and non-ferrous metals are used to adapt to industry tasks. At a machinery manufacturing enterprise, pre-sales and after-sales agents are used to understand and locate equipment error codes, helping frontline personnel handle debugging and testing issues.

Lenovo architects and engineers are stationed on site to identify requirements, debug, and release agents, while the Qingtian Innovation Lab in the back office continuously iterates industrial scenario capabilities. Enterprises' process, equipment, and business knowledge enter the development process, and the on-site team and the lab jointly follow up on site requirements and subsequent improvements.

For long-term management of model invocation, Lenovo's Token Hub provides unified access, task routing, usage aggregation, and permission control. Enterprises can view invocation consumption by department, project, and model, and manage resources through routing and cache configurations, bringing model access and usage into daily operations.

Moonshot AI delivers task outcomes such as code and documents with Kimi

Moonshot AI's Kimi invokes tools around user tasks, organizes agents, and delivers code or office documents. Foundation model R&D forms the core of this business; model capabilities reach users' work through products, tools, and services, and are also used in the team's own R&D.

In Kimi's Agent Swarm mode, after a user submits a task, the main agent organizes sub-agents to divide the work, invokes tools such as search and code, and aggregates the task results. Users can view the task list, tool execution steps, information sources accessed, and the progress of each agent, and ultimately obtain file outcomes such as website code, data analysis spreadsheets, or presentation slides.

During batch retrieval, relatively independent searches can be assigned to different sub-agents and then aggregated by the main agent. Different subtasks each retain relevant context and return key results to the main agent. The main agent integrates the research results and files accordingly, and users can continue to preview, download, or share the outcomes.

Model R&D also has its own usage records. In the technical introduction to Kimi K3, Moonshot AI presents GPU kernel optimization as an engineering task: in the later stages of K3 R&D, an early version took on most of the team's kernel optimization work. What the model participated in optimizing was the low-level code supporting computational performance, and researchers carried out engineering improvements based on it.

The training engineering that accompanies model capabilities also includes areas such as communication, compute operators, and agent environments. Moonshot AI has open-sourced related technologies, among which AgentEnv was developed by Moonshot AI in collaboration with KVCache.ai to provide a runtime environment for agent training. The training system complements model R&D and supports subsequent capability iterations.

Kimi K3 also treats attention mechanisms and cross-layer information transfer as architectural improvements, combining data and training methods to serve long-horizon coding, knowledge work, and reasoning tasks.

These R&D results are provided to users through the Kimi web product, Kimi Code, and API services. On October 23, K3.1 Pro launched in Kimi Code; on October 28, the desktop and command-line products continued to be updated, with model capabilities entering user tasks through continuously updated products.

DeepSeek uses its own models for internal agentic programming

When DeepSeek released the V4 preview, it listed using its own models for internal agentic programming as a practical application. Foundation model training, optimization, and inference constitute the core work of DeepSeek's user-facing services.

During the model training phase, DeepSeek builds foundational capabilities through large-scale data pre-training, then optimizes them with methods such as supervised fine-tuning and reinforcement learning to adapt the models to instruction and task needs. Pre-training data includes publicly available internet materials that have been acquired and filtered, as well as licensed data obtained through cooperation with third-party data providers. After deployment, the model performs inference based on user input, outputs text, tables, and code, and enters downstream applications through APIs.

The V4 preview released on April 24 has already been used for DeepSeek's internal agentic programming work. The model is adapted for tasks such as code and documents, completing longer execution processes through tools and agent frameworks. The team uses its own models for R&D, and the models' coding capabilities are also applied in internal engineering tasks.

External services are delivered through entry points such as the website, App, and API. Developers can invoke the model through APIs to use generation and inference capabilities in internal tools or user-facing applications; open model weights provide another deployment path for developers to integrate the model into specific tasks.

DeepSeek has also opened to developers search and web-fetching interfaces backed by its self-developed search engine. Developers can invoke retrieval tools to obtain the information needed for tasks, then combine it with the model's generation and inference services to organize the application's execution process.

On October 17, DeepSeek V4.1 Flash completed an update, and the official website, App, and API adopted the corresponding version in sync. DeepSeek continuously updates model capabilities and provides services through products and APIs, with R&D and service delivery centered on models.

Lenovo China reorganizes R&D roles, clarifies agent business responsibilities

Lenovo China has organized multiple roles in software R&D into a collaborative team of "full-stack engineers + five agents." Work such as requirements, architecture, front-end and back-end, testing, and operations is redistributed by role within the team; employees coordinate tasks, and agents take on the corresponding execution work.

As the human-machine division of labor changes, Lenovo China is also building positions such as forward-deployed engineers, business scenario experts, AI governance experts, and AI interaction designers, which respectively take on business understanding, engineering development, and operations management, while the responsibilities of full-stack engineers extend to cross-stage collaboration.

In job reshaping, Lenovo China distinguishes changes such as AI-enhanced, AI-extended, AI-disrupted, and AI-emerging: some positions retain their main body while handing part of their roles to agents, some extend professional capabilities through agents, and personnel further take on work such as knowledge accumulation and behavior management.

Sustained operation requires concrete management methods. Lenovo China has established a three-tier benefits management system of scenarios, projects, and program portfolios around revenue growth, cost savings, and efficiency improvement, linking agent development with business goals. On September 20, the World-Class Enterprise Research Institute of Renmin University of China, the Business School of Renmin University of China, and Lenovo Group jointly released the "AI-Native Organization Transformation Guide," summarizing related transformation practices.

In its organizational practice in China, Lenovo assigns business owners to agents. Goals are set by people, permissions are granted by people, and key decisions, exception handling, and result acceptance all have responsible persons, with records retained for traceability. Business owners follow up on agents' execution results, handle exceptions, and complete acceptance.

Lenovo is a global technology company founded in China, with revenue of RMB 589.9 billion in fiscal year 2025/26, and AI-related business revenue accounting for about 33% of full-year revenue; in the first quarter of fiscal year 2026/27, this share reached 35%. Lenovo has more than 20 R&D sites worldwide, and its global supply chain covers more than 30 manufacturing sites in 11 markets, with business covering 180 markets. In the first quarter of fiscal year 2026/27, Lenovo ranked first globally in market share for both PCs and AI PCs; in the second quarter of 2026, its x86 server shipments rose to first globally for the first time. Lenovo ranked first in mainland China and 51st globally on TIME's 2026 World's Best Companies list, and 153rd globally in the Fortune 2026 Global 500. Gartner's complete 2026 Global Supply Chain Top 25 list shows Lenovo ranked fifth globally, the highest among Chinese companies on the list. Lenovo was included in Forbes' 2026 lists of the World's Best Employers and the World's Top Companies for Women. Lenovo also received MSCI ESG's highest AAA rating, CDP's double A rating for climate change and water security plus an A rating for supplier engagement, and the EcoVadis Platinum Medal, and was included in Fortune's 2026 China ESG Influence List and S&P Global's Sustainability Yearbook (China Edition) 2026. In September 2026, Lenovo was awarded the Hang Seng Sustainability AAA rating, with its overall ESG score ranking first in the technology industry.