On August 3, Alibaba released its flagship model Qwen3.8 and simultaneously launched the public preview of Qwen Office. Qwen3.8-Max features 2.4 trillion parameters, with 95 billion activated parameters, and supports a context length of up to 1 million tokens. The domestic API pricing is set at RMB 12 per million tokens for input and RMB 36 per million tokens for output. Qwen Office integrates desktop, cloud, and enterprise collaborative agents, with future plans to connect with DingTalk IM, enterprise databases, and workflows. On the same day, Alibaba’s Hong Kong-listed shares rose over 6% during trading, reflecting positive market sentiment toward AI investments translating into revenue. The domestic office software market is highly competitive, with WPS, Tencent, ByteDance, and Alibaba each holding distinct advantages; the industry is now shifting from model capability competition toward application-specific scenarios, with enterprise adoption volume becoming a key metric for measuring the return on AI investment.

Article author, source: The Wall Street Journal

On August 3, Alibaba released its flagship model Qwen3.8, and "Qwen Office" entered public beta.

The model and office product have been launched in sync, completing a crucial piece of Alibaba’s AI narrative: Qwen3.8 delivers more advanced and cost-effective models, while Qwen Office integrates these models into high-frequency workflows.

The core of this round of trading lies in whether AI investments can be converted into revenue. Today, Alibaba's Hong Kong-listed shares rose over 6% during trading, receiving a positive response from the capital market.

Alibaba's valuation has been influenced by two competing narratives: e-commerce continues to generate cash flow but is impacted by slowing consumer growth, platform subsidies, and investments in immediate retail; AI and cloud computing offer higher growth expectations, but it remains to be seen when capital expenditures in these areas will translate into profits, requiring further performance validation.

If Qwen Office achieves enterprise-level adoption, Alibaba could introduce another product that drives token circulation.

Tech giants compete for the office market

Over the past two years, office agents have primarily focused on solving individual productivity issues—an employee can use an agent to search for information, write reports, and create spreadsheets, thereby reducing their own delivery time. However, corporate efficiency depends on handoffs between multiple departments: Can marketing data be integrated into the sales system? Can sales commitments be synchronized with legal and delivery teams? Can financial standards remain consistent across different teams?

Even if one stage is accelerated, if the data remains fragmented, permissions are not interconnected, and downstream teams must re-verify information, the organization’s overall delivery cycle may not be shortened.

The enterprise office market therefore requires a different product form. The product must understand organizational structure and permissions, connect to databases, knowledge bases, and business systems, support cross-departmental workflow orchestration, and maintain audit trails, version history, and accountability records. While individual agents focus on the quality of completing a single task, enterprise-grade products must also handle data governance, standard reuse, and multi-user collaboration. Ultimately, organizations purchase a manageable, replicable production system.

Competition in the domestic office market has also expanded.

WPS is advancing from its document entry point toward WPS 365 and the native office agent "WPS Lingxi," leveraging its strengths in formatting capabilities, existing document volume, and enterprise and government clients. Tencent is entering through the desktop agent, with WorkBuddy able to read and write local files, autonomously break down tasks, and invoke tools; the enterprise version can also connect to internal systems. Enterprise WeChat, QQ, Tencent Docs, and Tencent Cloud provide distribution channels and infrastructure for it.

ByteDance is integrating the capabilities of Doubao and Feishu to launch Doubao Enterprise, designed for organizations. The Seed model handles complex tasks, data processing, multimedia generation, and computer and browser operations, while Feishu provides enterprise knowledge, organizational permissions, and a collaborative environment. Documents, spreadsheets, and other AI-generated outputs can be directly integrated into Feishu workflows, enabling ByteDance to extend personal AI usage into team collaboration.

Ali's approach more closely resembles a comprehensive enterprise Agent architecture. Qwen Office integrates QoderWork, MuleRun, and Wukong, supporting desktop Agents, cloud Agents, and enterprise collaboration Agents, with future plans to connect to DingTalk IM, enterprise databases, and workflows. Qwen determines the intelligence ceiling, DingTalk provides organizational relationships, permissions, and process entry points, and Alibaba Cloud handles data, computing power, and token billing.

This full-stack combination represents Alibaba’s relatively clear advantage and tests whether the three systems can deliver a unified experience.

Law firms can institutionalize the M&A due diligence process as an organizational skill, and multinational teams can enable agents to continue advancing projects after employees have logged off. Such scenarios transform individual expertise into corporate assets, and once deployed in production, they consume far more tokens than a single Q&A or the generation of a piece of copy.

From stronger models to more token calls

Qwen3.8 provides a new model foundation for this entry. According to Alibaba’s disclosure, Qwen3.8-Max has a total of 2.4 trillion parameters, with 95 billion activated parameters, supports up to 1 million token context length, and significantly enhances capabilities in coding, coworking, and long-term agent tasks. The domestic API pricing is 12 RMB per million tokens for input and 36 RMB per million tokens for output, with a cached response price of 1.5 RMB.

Model capabilities determine whether an Agent can accomplish complex tasks, while unit cost determines whether a company is willing to scale usage. When both improve simultaneously, a typical demand elasticity emerges: as the price per invocation decreases, more tasks become economically viable; Agents expand from generating a single PowerPoint slide to continuously processing hundreds of documents, invoking multiple systems, and iteratively refining outputs—resulting in higher total token consumption per task.

Office scenarios can provide more stable and higher-volume query loads compared to consumer-facing Q&A.

Consumer-facing interactions typically consist of short dialogues, where user willingness to pay and retention can be highly volatile. Enterprise tasks, by contrast, often involve long documents, images, videos, and historical knowledge, and include multi-round cycles such as planning, tool invocation, and result validation. The workload for tasks like due diligence, auditing, or business analysis can far exceed that of ordinary chats.

Once a company integrates an Agent into its standard processes, the migration cost increases along with permission configurations, knowledge base accumulation, and skill development, and the quality of revenue typically surpasses that of one-time traffic conversion.

The agent's contribution to the cloud provider's performance has gradually become evident: for the quarter ending March 2026, Alibaba Cloud's revenue increased by 38% year-over-year, with external commercial revenue growing by 40%; revenue from AI-related products has achieved triple-digit growth for the 11th consecutive quarter, accounting for 30% of external cloud revenue, and the number of BaiLian customers has increased eightfold year-over-year.

Brokers are largely aligned in their assessment of this development: Citi defines Alibaba as a primary beneficiary of China’s token economy and estimates that MaaS could become Alibaba Cloud’s largest revenue-generating product; Morgan Stanley forecasts a 45% year-over-year increase in cloud revenue for the first quarter of fiscal year 2027, with cloud EBITA margin rising from 9% in the prior quarter to 11%; HSBC believes that the growing share of MaaS, expanded deployment of self-developed chips, and price increases in cloud products will collectively improve cloud profitability.

These changes reflect AI's core impact on Alibaba's valuation: the market is now measuring the return on Alibaba's AI investments using metrics such as token usage, MaaS revenue, and cloud profit margins.

Qwen3.8 delivers near-state-of-the-art capabilities at a lower price, which may temporarily reduce revenue per token; however, the lower cost will also encourage customers to delegate more processes to agents. As long as usage volume grows faster than price declines, and continued advancements in self-developed chips, caching, and sparse architectures further reduce inference costs, both cloud revenue and profit margins can improve simultaneously.

The market needs expansion of AI applications.

As leading models continue to improve in capability and inference costs keep falling, competition in the AI industry is extending toward real-world applications. Model rankings determine the technical ceiling, while real-world scenarios determine usage frequency, customer retention, and revenue scale.

The challenges in scaling scenarios are centered on enterprise workflows. Enterprises will ask whether the agent can reliably complete long-term tasks, inherit organizational permissions and protect data, integrate with existing systems while maintaining audit trails, and quantify how much human effort and time each task saves.

Therefore, the industry is seeking high-frequency, long-context, repeatable tasks. Programming, office work, customer service, marketing, and professional research have become primary entry points, with office work covering more departments and more complex data relationships. Vendors must simultaneously address model performance, system integration, and organizational governance, causing product formats to evolve from standalone assistants toward enterprise-grade execution platforms.

To determine whether the market judgment scenario is viable, a set of more concrete metrics is required: number of enterprise customers and payment rate, per-customer token consumption and retention, task completion rate, frequency of organizational skill reuse, and revenue and profit generated from usage volume.

For Alibaba, Qwen Office serves to integrate Qwen into enterprise workflows and convert tasks into cloud service usage on Alibaba Cloud. For the industry as a whole, the next phase of the AI narrative will emerge from more real-world scenarios and their ability to sustain continuous token flow.

Source: WeChat official account "Hard AI"