Unlocking AI Productivity: Inside JD’s Large‑Model Tool Innovations

Large Model Tool Development Status

After ChatGPT, large‑model research achieved breakthrough progress and is now widely used for text generation, intelligent customer service and other fields. Leading companies adopt large‑model tools, which are generally designed around five principles: low‑code UI interaction, security through local deployment, usability via multi‑end compatibility, scalability to follow open‑source trends, and high cost‑effectiveness for sustainable business.

JD’s Large Model Tool Capabilities and Applications

JD’s JoyAI spans models from 3B to 750B parameters, covering language, speech, image and other modalities. By improving inference efficiency and reducing training cost, JoyAI serves retail, logistics, health, industry and over one million merchants, powering agents such as JoyAgent, OxyGent and JoyCode.

JoyAgent

JoyAgent is the industry’s first 100% open‑source enterprise‑level intelligent agent, including front‑end, back‑end, framework, engine and core sub‑agents. It is fully decoupled from the platform, can be deployed locally, and enables high‑completion multi‑agent collaboration to accelerate AI deployment in production scenarios.

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

OxyGent

OxyGent is a multi‑agent collaboration framework that abstracts tools, models and agents as plug‑in operators (Oxy), allowing developers to compose multi‑agent systems like building blocks, offering extreme extensibility and end‑to‑end decision traceability.

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

JoyCode

JoyCode has evolved from early AI‑assisted coding (code completion / review) to a multi‑agent autonomous programming stage. Two successful cases are presented: front‑end H5 development using Figma MCP + JoyCode, and a custom agent that generates SDK integration code. To address target drift and large scenario variance, JoyCode 3.0 adopts a “plan‑then‑execute” strategy and custom agents, representing a new programming paradigm.

Experience and Business Practices

The three key factors for large‑model application are selecting the right tool/platform, defining quantifiable goals, and iterative phased validation. JD’s practice starts with MVP verification (e.g., low‑code workflow), then continuously optimizes through a data‑closed loop (e.g., accuracy, fluency metrics). Good tools reduce effort on model stability, permission security and multi‑system integration.

Product managers note that AI applications must accept gradual effect improvement, using gray‑scale rollout and scenario focus to close the value loop, and predict that agents will evolve toward multimodal interaction and embodied forms such as digital employees.

Office Scenario

AI‑enabled office tools have formed a systematic methodology at JD, focusing on intelligent utilities like unread summary and automatic meeting minutes, and launching the Super Assistant Max that integrates tasks and information flow, enabling “one‑sentence task creation”. Future work emphasizes “digital employees” with thinking, execution and collaboration abilities, building an open ecosystem for custom AI assistants.

Code Review

JD processes nearly ten thousand code reviews daily. To break the efficiency bottleneck of manual review, the AI review system built on JoyAI, DeepSeek and AutoBots performs line‑level detection covering standards, security vulnerabilities, and integrates 17 rule libraries into the Coding platform, reducing Java defect rates by 32% for the logistics team. The system now supports multi‑language access across the group and is tightly integrated with JD’s office platform for automatic review progress notifications.

Local Life

Three typical directions for large‑model innovation in local‑life business are: cross‑language code translation (bridging Python with Java/Go), general knowledge labeling to resolve regional knowledge conflicts, and a multi‑agent simulation system that models rider, merchant and user dynamics to optimize dispatch strategies and lower trial‑and‑error costs.

Logistics

JD Logistics has successfully applied AI in several areas: digital warehouse AR glasses using multimodal perception (vision + AR) improve picking efficiency by 15%; an intelligent outbound call system automatically generates customer‑facing language; an address parsing engine based on JoyAI achieves building‑level precision, reducing mis‑delivery rate to below 0.00015% and dramatically lowering complaint rates; and an elevator recognition model determines elevator presence and usage, saving rider time.

Future Outlook of Large Model Tool Construction

Policy: The rapid development of large‑model technology must align with regulatory environments. Recent Chinese “Interim Measures for Generative AI Services” and the EU AI Act emphasize data quality, content safety, copyright protection and systematic risk management, urging enterprises to establish strict data cleaning, sensitive‑information filtering, and compliance mechanisms.

Agent’s ultimate form: agents will evolve toward multi‑agent collaboration for complex tasks, embodiment as digital humans (2D/3D) and physical robots, and collective intelligence that continuously updates knowledge bases and decision logic through environment interaction, supporting dynamic team composition.

Embodied intelligence end‑to‑end path: through multimodal perception fusion, data closed‑loop learning, and lightweight deployment (e.g., INT4 quantization, NPU/TPU acceleration), large models can replace traditional modular rule‑based pipelines with holistic perception‑decision‑control optimization, enabling systems like autonomous driving to handle complex scenarios intuitively.