From Ant Ling's New Model to JD 11.11: AI Deployment Race Focuses on Task Delivery and Scenario Closed Loops

AI is moving from 'chatting' to 'doing work'. In mid-October, Ant Ling's new open-source model, the rollout of AI scenarios for JD 11.11, Apple Intelligence's integration with Alibaba Qwen, plus reports on enterprise Agent deployment and device-side music co-creation all pointed, almost simultaneously, to the same shift: beyond model scores, task delivery, system closed loops and real-world scenario fit are becoming the focus of competition.^[1][2][4][6]

Ling-3.1-flash from Ant Ling: Packing Long Tasks into a Hybrid Architecture

Ant Ling has open-sourced Ling-3.1-flash, officially positioned as a model that 'gets the work in real workflows done'. It has about 560B total parameters, with about 25B activated per token; its attention structure continues that of Ling-3.0-flash, increases the proportion of linear attention, and uses 7 layers of KDA with 1 layer of Gated MLA, retaining 1 layer of Gated MLA every 8 layers to maintain a global attention pathway. Its context window can reach up to 1M. It sets 512 routed experts, selects 8 of them per token, and simultaneously calls 1 shared expert. The team also used mixed-precision quantization to reduce model weight footprint from about 534GB to 292GB; in the corresponding Hopper test configuration, the GPUs required for deployment dropped from 8 cards to 4 cards.^[1]

In the office task benchmark GDPVal-AA V2.1, Ling-3.1-flash scored 1673 ELO; according to the latest public evaluation by Artificial Analysis, its intelligence index is 41, output speed is about 210 tokens per second, and blended API price is about $0.19 per million tokens. In the HealthBench Professional evaluation for healthcare scenarios, it also maintained a relative lead. A better example of its long-task capability is an internal Ant long-duration Lua native compiler development task: the whole process lasted about 17 hours, during which the model repeatedly developed and corrected, and the final independent test pass rate was 97.8%. It has been connected to Ant's internal desktop collaborative Agent client 'Xiaohu', where the model directs actions and the execution environment reads files, runs programs and writes results.^[1]

JD 11.11: AI Enters More Than 3,000 Scenarios

JD 11.11 fully kicked off at 8 p.m. on October 12, with the number of first-day ordering users achieving double-digit year-on-year growth. On October 15, at the 2026 11.11 'Better and Cheaper' press conference, JD announced a comprehensive service upgrade, showcased for the first time more than 100 segmented service products, and released the JD Service '333' plan. AI became a key variable in this year's shopping festival: AI shopping assistant 'Dongdong', JoyInside embodied intelligence, JD digital human JoyAvatar, and the 'Super Brain + Wolf Pack' robot army and other AI results have been fully applied in more than 3,000 scenarios including retail, logistics, health, industry, food delivery and housekeeping.^[2]

For merchants, JD Jingmai AI Business Center opened dozens of core AI tools, covering the complete business chain of product selection, listing, operations, design, fulfillment and after-sales; JD digital human JoyAvatar, with round-the-clock live streaming, became a normalized service engine for the content ecosystem of more than 90,000 merchants during 11.11; JD Bailing digital employees launched marketing digital employees to help merchants independently generate marketing strategies and execute them with one click. On the service side, the JD Service Panorama made its debut, bringing together more than 100 segmented service products; the scale of productive services has exceeded RMB 300 billion, serving more than 10 million enterprise customers, covering 91% of Fortune Global 500 companies in China, 98% of China's top 500 private enterprises and 70% of specialized and sophisticated enterprises.^[2]

Apple Intelligence Integrates Alibaba Qwen: Test Version to Launch in Late October

On October 15, during a visit to Shanghai, Apple Executive Chairman of the Board Tim Cook said that Apple Intelligence and the related Alibaba Qwen large model have indeed been approved, and a test version will be released later in October. It will include the first batch of Apple Intelligence features such as real-time translation, writing tools and photo clean-up tools, but Siri AI will not be included this time. Cook said: 'The first batch of core Apple Intelligence features can be rolled out slowly.' This progress means Apple's AI features in the Chinese market will be deployed with the help of the Alibaba Qwen large model. The first batch of features chooses relatively mature, risk-controllable capabilities such as translation, writing and photo cleanup, rather than directly launching the voice assistant.^[4]

The same news also mentioned that Alibaba Cloud and Nvidia reached cooperation in the field of Physical AI. For Apple, the localization of Apple Intelligence is not just a matter of model integration; it also involves regulatory approval, rolling out features in batches and the pace of user experience; launching core features first in a test version and then gradually expanding is a more prudent path.^[4]

Enterprise Agent Deployment: Six Major Breakpoints and Full-Chain Closed Loops

QbitAI Think Tank released the report 'Agent-Driven Enterprise Intelligence: From Point Efficiency Gains to Large-Scale Reinvention', attempting to answer why enterprise AI often remains in a state where 'it works locally but cannot move as a whole'. The report strings together the problem through a store out-of-stock risk response scenario: the technical team thinks the report is done once it is generated, the business staff think the matter is accounted for only when the cause is clarified, and managers care about whether the stockouts actually decreased; the three parties' standards for completion are not consistent. At the data level, there is a difference between book inventory and the amount that can truly be adjusted; if AI makes judgments using book numbers, its starting point is skewed. Even if the analysis is correct, between recommendation and execution there are still links such as manual form filling, submitting approvals and tracking delivery.^[6]

The report summarizes these sticking points into six major breakpoints: task definition, data credibility, verifiability of judgment, action implementation, division of authority and responsibility, and value measurement and replication. In response, the report proposes that a mature enterprise-grade Agent must have full-chain capabilities that connect tasks, context, judgment, action, control and measurement. In terms of cases, Sanhua Intelligent Controls used Shufan to unify data and semantic definitions at the underlying layer, allowing AI to automatically pull budget-versus-actual data and identify deviations, compressing budget analysis that originally took 2-3 hours to 10 minutes; Kongge Digital Technology shortened the development cycle of a new energy CRM project from 60 person-days to 16 person-days; Shanghai Qitong Digital Technology achieved about 2.5x efficiency improvement in BIMS system development, with a first-time feature availability rate of 78.63%.^[6]

Device and Scenario Co-Creation: STEPX Neo Lets Listeners Become Music Co-Creators

From October 17 to 18, the top electronic music festival The Magic of Tomorrowland 2026 was held in Shanghai. As a strategic partner, STEPX brought its large-model-native agent phone STEPX Neo to the venue, exploring the integration of AI, music and trend culture, and unlocking the agent gameplay of 'everyone can create their own AI music'. STEPX Neo, officially released on October 13, runs the agent-native operating system Step AOS and comes with the new-generation personal agent Step Amoo; STEPX and AI makers created an on-site music co-creation experience based on the StepAudio 3 Music model and the STEPX Neo agent phone, allowing the audience to turn personal inspiration into exclusive electronic music and move from listeners to creators.^[3]

STEPX also launched Step Studio, its first AI terminal-native development tool. Makers only need to use natural language to generate AI apps, games, agents and Skills. Based on this development environment, STEPX Neo can fully open model, system, hardware and ecosystem capabilities to makers, and supports connecting external services and professional agents through protocols such as MCP and A2A. The exclusive AI music gameplay is exactly the deployment of this open capability in a young trend culture scenario, showing how an agent terminal can receive inspiration and amplify personal imagination and creativity.^[3]

Business Schools Respond to the AI Era: Training People Who Can Define Problems

As AI can quickly process massive information and output professional analysis, some traditional business school teaching methods focused on knowledge acquisition, case analysis and standardized problem-solving are facing disruption. Antai College of Economics and Management, Shanghai Jiao Tong University released the 'Antai Manifesto for Business Talent Development in the AI Era', systematically responding to the challenges AI brings to business education. The direction given by SJTU Antai MBA is to move further from imparting knowledge and problem-solving methods to cultivating the ability to define problems, think independently, integrate across disciplines and make value judgments.^[5]

On August 16, SJTU Antai MBA officially released the 'Five-Chain Talent' development system for the intelligent era: linking intelligence, linking science and innovation, linking industry, linking the world, and linking with SJTU. According to the plan, over the next decade, SJTU Antai MBA will work with companies across industries to jointly create and implement 1,000 exclusive AI real-world projects. On September 6, it announced the first batch of AI scenario lists and co-building ecosystem partners for its Hundred Industries, Thousand Enterprises initiative, and reached a strategic cooperation with the AI open-source learning community Datawhale, simultaneously launching the FDE talent specialized training program and releasing 'AI+X' practical training courses and selected FDE courses. The development system highlights three practical capabilities: AI4O, FDE and OPC. For the 2027 cohort, the program length will be uniformly adjusted to 2 years, implementing a flexible 40-50 credit system, with practical courses accounting for no less than one quarter of total class hours.^[5]

Common Trend: From Model Scores to Task Delivery

Putting these clues together, the focus of AI competition is shifting from model parameters and leaderboards to delivery capability and system closed loops in real tasks. Ant Ling's Ling-3.1-flash uses hybrid linear plus sparse MoE to lower the cost of long tasks, and is connected to execution environments such as 'Xiaohu'; JD spreads AI into more than 3,000 scenarios including retail, logistics, health and industry, while using its service system and digital employees to complete the business chain; Apple chose to integrate Alibaba Qwen and roll out core features in batches via a test version; the enterprise Agent report points out that breakpoints such as data credibility, authority and responsibility, and action feedback are harder to overcome than model capability.^[1][2][4][6]

Going forward, several directions are worth watching: the actual experience and pace of feature expansion after the Apple Intelligence test version launches; real data on how AI tools perform in merchant operations and consumer shopping during JD 11.11; whether enterprise Agents can move from closed loops in a single high-value scenario to reusable metric definitions, interface components and approval rules; and whether SJTU Antai MBA's AI real-world projects and FDE training program can bring industry needs into the classroom earlier. On the device side, STEPX Neo's music co-creation offers a window into how young people create together with agents.^[3][5][6]