JD.com’s Cloud open-sources JoyAI model as “OpenClaw” rollout drives token usage up 455%
JD.com is pushing its artificial-intelligence stack from in-house productivity tool to sellable infrastructure, open-sourcing a 48B-parameter JoyAI model and packaging deployment products that the company says lifted JD Cloud token calls 455% in a week—an early signal that China’s enterprise AI spending is shifting from “model choice” to “implementation speed.”
The latest turn came late Tuesday in Beijing, when JD’s technical team unveiled the “OpenClaw” (“lobster”) line—one-click lightweight cloud host images, an all-in-one appliance and cloud SaaS—alongside what it called a “CodingPlan” large-model bundle. JD said OpenClaw’s cloud-service user base grew more than 300% week-on-week, indicating demand for turnkey rollout rather than bespoke model engineering.
Initial market implications center on whether JD Cloud can convert usage growth into durable enterprise contracts, as Chinese hyperscalers and platform firms compete to lock in workloads through deployment tooling, compliance features and ecosystem integration—not just benchmark scores.
Open-sourcing expands reach while tightening ecosystem lock-in
JD for the first time open-sourced the instruct version of JoyAI-LLM Flash, a general-purpose foundation model with 48B parameters and 3B “activated” parameters. JD said the model outperformed GLM-4.7 Flash (non-thinking) and other peers of similar scale in its published tests, while using a dynamic sparse routing approach intended to improve compute efficiency.
For investors, the open-source move reads less like a giveaway and more like a distribution strategy: by placing JoyAI-LLM Flash on Hugging Face under the jdopensource account, JD can seed developer adoption and pull inference and tooling demand back into JD Cloud, where it controls billing, observability and enterprise support. JD said its JoyAI technology is already used in more than 2,000 scenarios across its “super supply chain,” and that internal agent deployments have exceeded 50,000—figures that matter because they imply the company has a large internal reference base to productize.
Productizing deployment targets compliance-heavy enterprise budgets
JD Cloud said it has preloaded an OpenClaw application image into its lightweight cloud host product, enabling a three-step deployment so developers don’t need to build runtime environments manually. For larger companies, JD introduced an OpenClaw all-in-one appliance positioned around “zero-code” setup, native open-source ecosystem compatibility and official updates.
JD described three hardware configurations: a standard model aimed at data-security compliance that supports more than 80 concurrent users and processes more than 1 billion tokens per day; a second standard model targeting higher accuracy and concurrency with support for 50 concurrent users and more than 500 million tokens per day; and a personal version for teams of five or fewer that processes about 350 million tokens per day.
That lineup reflects a China-specific buying pattern: many enterprises want on-premises or hybrid deployments for data governance and sector regulation, while still demanding rapid iteration. By selling appliances and images rather than only APIs, JD Cloud is effectively competing for the same procurement lines as traditional infrastructure vendors—while using open-source compatibility to reduce switching friction.
Digital humans and embodied data shift competition to proprietary datasets
JD also released JoyAvatar, a self-developed digital-human video generation framework. The company said JoyAvatar surpassed Omnihuman-1.5 and KlingAvatar 2.0 on internal evaluations and can generate videos longer than 30 seconds while maintaining identity stability. It reported a lip-sync similarity score (Sync-C) of 5.57 and a hand keypoint confidence (HKC) of 0.87, and positioned the technology for ecommerce livestreaming, customer service and content creation.
The longer-term competitive edge JD emphasized is data. JD said it aims to become the world’s largest embodied-intelligence data company, targeting 5 million hours of real-world human-scene video data within one year and more than 10 million hours within two years, alongside 1 million hours of robot “body” data. If executed, that would move the moat from model architecture—where open-source compresses differentiation—toward proprietary, scenario-rich datasets that can tune robotics and multimodal systems for commerce and logistics.
Hardware platform integrations aim to monetize via devices and channels
JD said its JoyInside capabilities—built to adapt AI to hardware terminals—added social functions and upgraded voice synthesis in early 2026, and expanded interaction to eight dialects. The company said JoyInside has connected with nearly 100 home-appliance and home-furnishing brands and more than 40 robotics and AI-toy makers, and it launched a JoyInside developer platform offering a low-code environment plus modules and industry resources.
For JD, the monetization path is twofold: cloud usage driven by AI agents and digital humans, and device-side licensing plus retail-channel leverage for partners. The tighter the integration between models, deployment tooling and JD’s retail distribution, the harder it becomes for enterprise customers and device makers to treat AI as a pure commodity.
Related Coverage:
JD.com Q4 2025 Earnings: Operating Loss Driven by New Business Drag, Leverage Rises
JD.com Founder Liu Qiangdong Bets RMB 5 Billion on Yachts, Targeting China's Untapped Mass Market