How JD’s Large‑Model Tools are Shaping AI in Enterprise: Insights & Roadmap

With the rapid development of large‑model technology, AI tools have become essential productivity assistants, dramatically improving efficiency, solving business pain points, and extending knowledge through intelligent collaboration. Mastering these tools is now a key competitive advantage for modern professionals.

1. Industry Development Status of Large‑Model Tools

Since the emergence of ChatGPT, large models have made breakthrough progress and are widely used in text generation, intelligent customer service, and more. Leading companies follow five design principles:

Low‑code interaction (UI principle): Reduces the usage barrier so non‑technical users can quickly adopt.

Security (local deployment): Private deployment prevents data leakage.

Usability (multi‑device compatibility): One code works on many devices, covering scenarios comprehensively and lowering adaptation costs.

Scalability: Aligns with open‑source trends for continuous iteration.

Cost‑effectiveness: Ensures commercial sustainability.

2. JD’s Large‑Model Tool Capabilities and Applications

JD has upgraded its Yanxi large‑model brand to JoyAI , offering models from 3B to 750B covering language, speech, and image modalities. Innovations improve inference efficiency and reduce training costs. JoyAI serves retail, logistics, health, and industry, supporting over a million merchants and hundreds of business scenarios. JD has built an enterprise‑level AI Agent platform JoyAgent , an intelligent coding assistant JoyCode , and a multi‑agent collaboration framework OxyGent .

JoyAgent

JoyAgent is the industry’s first 100% open‑source enterprise‑level intelligent agent, providing front‑end, back‑end, framework, engine, and core sub‑agents. It can be deployed locally and enables rapid production‑scene deployment through high‑completion multi‑agent collaboration.

Open‑source address: https://github.com/jd-opensource/joyagent-jdgenie

OxyGent

OxyGent abstracts tools, models, and agents as plug‑in atomic operators (Oxy), allowing developers to assemble multi‑agent systems like building blocks, offering extreme extensibility and full‑traceability.

Open‑source address: https://github.com/jd-opensource/OxyGent

JoyCode

JoyCode has evolved from AI‑assisted coding (completion/review) to multi‑agent autonomous programming. Two success cases are highlighted: (1) Front‑end H5 development using Figma MCP + JoyCode; (2) Custom agent generating SDK integration code. JoyCode addresses challenges such as goal drift and scenario variance by “plan‑first‑execute” strategies and custom agents.

3. Experience and Business Practice

Three key factors for large‑model deployment: choose the right tool/platform, define quantifiable goals, and iterate in phases. JD’s practice starts with MVP validation (e.g., low‑code workflow), then closes the data loop for continuous optimization.

Office Scenario

AI‑enabled office tools include unread summarization, automatic meeting minutes, and the Super Assistant Max, which creates tasks from a single sentence. Future focus is on “digital employees” that think, execute, and collaborate, forming an open ecosystem for custom AI assistants.

Code Review

JD processes nearly ten thousand daily code reviews. Using Yanxi models, DeepSeek, and AutoBots, AI performs line‑level detection for standards, security, and integrates 17 rule libraries into the Coding platform, reducing Java defect rates by 32% and providing multi‑language support across the group.

Local Life Applications

Cross‑language code translation: Bridges Python algorithms with Java/Go engineering stacks.

General knowledge labeling: Resolves regional knowledge conflicts, replacing manual rule maintenance.

Multi‑agent simulation: Simulates rider‑merchant‑user dynamics to optimize delivery dispatch.

Logistics Applications

Digital warehouse AR glasses: Multi‑modal vision + AR prompts boost picking efficiency by 15%.

Intelligent outbound call system: Generates customer‑facing scripts for better service.

Address parsing engine: Building‑level location accuracy reduces mis‑delivery to <0.0015%.

Elevator recognition model: Detects elevator presence and usage, saving rider time.

4. Future Outlook

Policy Perspective

Regulations such as China’s “Interim Measures for Generative AI Services” and the EU AI Act shape data quality, content safety, and copyright protection. Enterprises must align with these frameworks to mitigate policy risks and seize innovation opportunities.

Agent End‑State

Agents will evolve toward multi‑agent collaboration, embodiment (digital humans, robots), and collective intelligence, enabling complex task handling, physical interaction, and self‑evolving ecosystems.

Embodied Intelligence Path

By integrating perception, data closed‑loops, and lightweight deployment (e.g., INT4 quantization), end‑to‑end models can replace rule‑based pipelines, achieving “intuitive” handling of complex scenarios such as autonomous driving.

Marketing Opportunities in the AI Era

AI‑driven conversational advertising will become mainstream, leveraging new traffic entry points, precise vertical recommendations, and breakthroughs in intent recognition and generative ranking to transform traditional ad logic.

JD’s technology team invites peers to embrace the AI wave and co‑create the next efficiency revolution.