--- title: "JD rebuilds its AI around JoyAI — a 3B-to-750B model family and a fully open agent" date: 2026-10-07 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-jd-joyai-full-size-family language: en --- # JD rebuilds its AI around JoyAI — a 3B-to-750B model family and a fully open agent > At WAIC 2025 JD upgraded its Yanxi (言犀) brand to JoyAI, a full-size model family from 3B to 750B, and open-sourced its enterprise agent JoyAgent. JD.com is best known as the warehouse-and-delivery rival to Alibaba. But at the World Artificial Intelligence Conference (WAIC) in Shanghai on July 26, 2025, the retailer made a different kind of announcement: it upgraded its Yanxi (言犀) large model (大模型) brand to **JoyAI**, and revealed a model family spanning from a tiny 3B model to a 750B flagship — covering language, speech, image, video and digital-human avatars. ## A model for every shelf The headline is the size range. Instead of one flagship, JD now ships a ladder: small models that run close to the user (in a store kiosk, a logistics scanner, a toy), and a 750B model for heavy reasoning. The pitch is that one coordinated family can serve a huge retailer's every touchpoint — customer chat, product copy, warehouse QA, doctor scheduling in its healthcare arm — without stitching together models from a dozen vendors. JD says the family leans on "dynamic hierarchical distillation" and "cross-domain data governance" to keep the small models competent. The claimed payoff is concrete: average inference efficiency up 30%, training cost down 70%. Those are vendor figures, but the direction is the same efficiency story repeated across Chinese labs in 2025 — squeeze more capability out of less compute. ## An agent (智能体) you can actually deploy The more interesting release at WAIC was JoyAgent, which JD Cloud describes as the industry's first 100% open-source enterprise agent (智能体). "100% open" means the front end, back end, framework, engine and core sub-agents are all published — a developer can stand up a private, enterprise-grade multi-agent product without licensing a platform. JD reports the GitHub project passed 1,000 stars within three days of release. JoyAgent is built around a multi-agent collaboration engine and a dynamic DAG execution engine — essentially a scheduler that breaks a complex business task into steps and runs them across models of different sizes. A 2.0 version had arrived in May 2025 on a mixture-of-agents design with multimodal input; a 3.0 followed in September 2025, adding self-coding, multimodal retrieval-augmented generation (RAG), and a DataAgent, with the DCP data-governance module also open-sourced. ## Why a retailer owns a model factory JD's logic mirrors Meituan's and Alibaba's: if AI is going to reroute how people shop, the company that controls the model controls the experience. JD's edge is its supply chain — it knows inventory, logistics, returns and pricing at a granularity no general model does. A JoyAI-powered assistant can, in principle, not just recommend a product but check whether it is in stock, how fast it ships, and what the return policy is, in one breath. JD also launched JoyInside, an "embodied intelligence" brand that pipes the JoyAI model into robots and AI toys, giving them conversational interaction. That is the retail giant's answer to the humanoid (人形机器人) wave: put the model inside the things it already sells and ships. ## Reading the trend According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, JD's move shows how Chinese internet giants are converging on the same playbook — a full-size open model family plus a fully open agent framework — because owning both the model and the deployment layer is what protects a platform from being displaced by someone else's AI. The throughline across JD, Meituan and the model labs is clear: in China in 2025, "having an AI strategy" increasingly means "shipping your own weights." ## Honest limitations The JoyAI size range, the 30% / 70% efficiency figures, and the JoyAgent open-source claims come from JD's WAIC presentations as reported by China News Service and state radio (CNR), plus JD Cloud's own materials. The 30% and 70% numbers are vendor-stated and not independently audited; real gains will vary by workload and hardware. "100% open-source" refers to the agent framework's code, not necessarily to the underlying 750B model weights, which JD has not broadly open-released. Benchmark comparisons against Western models were not part of the launch, so claims of "first-tier" performance for the 750B model should be treated as the company's positioning rather than verified ranking. As with all platform models, expect commercial framing in sample outputs. ## What readers can do now - If you run enterprise workflows, JoyAgent's fully open code on GitHub is a rare chance to self-host a multi-agent system without a vendor lock-in contract — pilot it for internal ops before buying a closed platform. - For China retail, logistics or healthcare deployments, evaluate JoyAI against the specific supply-chain advantages JD emphasizes; the differentiator is data access, not raw model size. - Don't assume the 750B weights are open just because the agent is — verify licensing before any commercial use. - Watch for the JoyInside embodied line if you build robots or smart devices; that is where JD's model-meets-hardware bet will show up first. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-jd-joyai-full-size-family Free to quote with attribution and a link to the original.