JD.com Announces AI Progress: "Lobster" Series Products Show 455% Week-over-Week Growth in Token Calls

On March 24, JD.com announced partial progress in AI research and development, including open-sourcing models, digital humans, embodied intelligence, and “Lobster.”

In terms of large models, JD.com for the first time open-sourced the instruct version of its foundational large model JoyAI-LLM Flash, stating that this model is suitable for code development, agent building, terminal applications, and other fields, especially compatible with the recently popular “Lobster.”

JD Cloud said that to address developers’ concerns about the “Lobster” agent and the high token costs behind it, they have launched multiple product forms based on the JoyAI large model, including lightweight cloud hosts with one-click deployment, all-in-one machines, and cloud SaaS versions through open-sourcing OpenClaw architecture. According to statistics, after the launch of JD Cloud’s “Lobster” series, token call volume increased by up to 455% in just one week.

In addition to “Lobster” related products, JD.com has updated its digital human technology again. JD Digital Human JoyStreamer uses dual-teacher DMD training technology, dynamic CFG modulation strategies, and techniques like historical frames plus pseudo-last frame structures to solve industry issues such as audio-video desynchronization, multimodal control inconsistency, and identity distortion in long videos. It also introduced a new “free state digital human” feature, targeting industries like home appliances, fashion, and apparel.

Over the past year, JD has made multiple deployments in digital humans. In December last year, JD announced that digital human live streaming would be freely available to all merchants. According to Yicai, cost reduction and efficiency gains are key reasons merchants choose digital human live streaming, with costs as low as one-tenth of real-person live streaming. By 2025, digital human live streaming is expected to generate hundreds of billions of yuan in GMV.

JD’s digital human and AIGC product leaders told Yicai that more than 70,000 merchants are currently using JD’s digital humans. Beyond cost savings, merchants’ requirements for digital humans are increasing, making them more like real hosts—for example, enhancing interactivity to boost viewer engagement. Improving these experiences requires not only technological support but also database updates, such as uploading more after-sales knowledge, enriching Q&A details, and improving interaction quality.

Additionally, some issues remain in digital human hosts in live streams, such as monotone voices and unoptimized clothing details. The responsible person explained that current technology can address these issues. The flaws viewers see are not due to technological lag but application lag—meaning the digital humans used by merchants have not been updated. One of the industry’s future directions is self-iterating technology, similar to the current “shrimp farming” concept.

From an industry perspective, digital human IPs are gaining attention. The key to creating digital human IPs is not just technology but whether it is worth developing the IP. Once the technical foundation is solid, creating IPs also requires investments in other areas, such as marketing and design.

Besides these developments, JD.com announced data collection goals in the field of embodied intelligence. In this area, JD launched JoyInside, an embodied intelligence system. So far, JoyInside has established deep collaborations with nearly 100 home appliance and furniture brands, over 40 robot and AI toy brands.

A major challenge in embodied intelligence is the lack of real-world data. During data collection, JD plans to leverage its advantages in retail, logistics, industry, food delivery, and domestic services to build the world’s largest and most comprehensive embodied intelligence data collection center. According to the plan, JD will mobilize over 100,000 internal employees across various professions and up to 500,000 external industry personnel, including over 100,000 citizens in Suqian, to carry out the “largest human-scale data collection campaign.”

Within the next year, JD aims to accumulate 5 million hours of real-world human scene videos, surpassing 10 million hours in two years, while also collecting 1 million hours of robot body data.

(Source: Yicai)