JD.com AI Solution: Open-Source JoyAI-LLM Flash Reduces Costs and Boosts Efficiency, Digital Human Performance Surpasses SOTA, Embodied Intelligence Covers Hundreds of Brands

AI technology is accelerating its iteration, but balancing cost efficiency with performance and bridging the “last mile” of industrial implementation remain common challenges for the industry. On March 24th, JD.com announced some phased progress in its AI R&D and application, proposing solutions to the aforementioned challenges.

01 Open-Source Large Model JoyAI-LLM Flash Significantly Lowers the Barrier to AI Use

At the model capability level, JD.com has open-sourced the Instruct version of its foundational large model, JoyAI-LLM Flash, for the first time. This model innovatively introduces the FiberPO reinforcement learning method based on “fiber bundle” theory, achieving a technological breakthrough. In 19 authoritative benchmark tests, the model performed excellently, reaching top-tier industry levels for models of comparable parameter scale.

The JoyAI-LLM Flash model possesses core advantages of efficient response, lightweight deployment, and high accuracy, enabling precise adaptation to various fields such as code development, agent construction, and terminal applications. Its design aims to provide developers and small-to-medium enterprises with a cost-effective model foundation, thereby significantly lowering the barrier to using AI technology.

02 JD Cloud Intelligent Agent Solution Sees 455% Weekly Growth in Token Calls

Addressing the high token costs behind intelligent agent applications, JD Cloud, based on the JoyAI large model and through the open-source OpenClaw architecture, has launched various product forms including one-click deployment on lightweight cloud hosts, all-in-one machines, and cloud-based SaaS versions, covering diverse needs from individual developers to medium and large enterprises. By optimizing inference costs with a self-developed cloud-native inference framework and launching the CodingPlan large model package, it continuously reduces usage costs.

According to statistics, after the launch of JD Cloud’s related intelligent agent series products, token calls in the past week have increased by a staggering 455% compared to the previous period.

03 JD Digital Human JoyStreamer Performance Surpasses International SOTA Levels

JD’s digital human JoyStreamer has achieved core technological breakthroughs. Through three innovative technologies, it addresses industry pain points such as audio-video desynchronization, uncoordinated multimodal control, and identity distortion in long videos. Its performance surpasses international SOTA levels, with core metrics reaching commercial-grade standards, adaptable to over 20 industry scenarios including e-commerce livestreaming, cultural tourism narration, and news anchoring.

• Dual-Teacher DMD Post-Training Technology: Enables the digital human to accurately respond to complex instructions and perform complex actions without requiring additional training data.
• Dynamic CFG Modulation Strategy: Fundamentally resolves multi-dimensional control conflicts in digital humans, such as “uncoordinated limb movements and facial expressions, mismatched lip-syncing and voice.”
• Historical Frame + Pseudo-Last Frame Structure: Can stably generate long videos over 30 seconds, ensuring the digital human’s appearance remains undistorted and movements are smooth throughout, meeting the needs of commercial scenarios like livestreaming and long-form videos.

Leveraging JD’s JoyAI large model and the aforementioned technological breakthroughs, JoyStreamer has launched the “Free-State Digital Human,” precisely adapted for five major industries including home appliances/furnishings and fashion/apparel. This digital human breaks the limitations of traditional digital humans, such as rigid movements and fixed postures. It supports natural walking, flexible posing, smooth camera following, and entering/exiting the frame, maintaining high-fidelity quality even when the face is partially obscured.

04 JoyInside Deeply Collaborates with Over a Hundred Brands, Equipping Everything with an AI Brain

In the field of embodied intelligence, JD.com, leveraging its multi-scenario intelligent interaction expertise and JoyAI large model capabilities, has launched the JoyInside Embodied Intelligence solution, aiming to implant a “high emotional intelligence brain” into devices like smart home appliances and robots.

JoyInside boasts advantages such as hyper-realistic dialogue, full age-range adaptation, and highly coordinated speech and actions. Currently, this solution has established deep collaborations with nearly a hundred home appliance/furnishings brands and over 40 robot and AI toy brands, covering diverse scenarios.

Furthermore, JD.com has created the JoyInside one-stop development platform, aiming to achieve full closed-loop empowerment for AI hardware from concept to mass production. JD.com will soon host the “AI Terminal New Species” Innovation Competition, inviting global developers, hardware manufacturers, and AI innovation teams to submit innovative AI hardware solutions based on JoyInside interaction. Outstanding projects will receive full-chain support from concept incubation to product distribution.

05 Building an Embodied Data Collection Center to Solve the Industry’s “Data Famine”

The scarcity of real-world scenario data is a major pain point in the current embodied intelligence field, constraining model training and industrial implementation. To solve the “data famine” at its source, JD.com is leveraging its rich advantages in retail, logistics, industrial, and other scenarios to build the world’s largest and most comprehensive embodied intelligence data collection center.

This data collection initiative will mobilize over 100,000 JD.com internal employees as well as external personnel from various industries. It plans to accumulate 5 million hours of real-world human scenario video data within the next year, exceeding 10 million hours within two years, while simultaneously collecting robot body data to propel the embodied intelligence industry from algorithmic simulation towards real data-driven development.

The above progress represents part of JD.com’s AI R&D and application achievements. Currently, JD.com has already facilitated the rapid implementation of the JoyAI large model in multiple scenarios including retail, logistics, health, and industry. In the future, JD.com stated it will continue with the combined strategy of “AI Infrastructure + Model Capability + Scenario Implementation,” working hand-in-hand with industry partners to promote the deep integration of AI and industry.

⚠️ หมายเหตุ: เนื้อหาได้รับการแปลโดย AI และตรวจสอบโดยมนุษย์ หากมีข้อผิดพลาดโปรดแจ้ง

☕ สนับสนุนค่ากาแฟทีมงาน

หากคุณชอบบทความนี้ สามารถสนับสนุนเราได้ผ่าน PromptPay

SCAN TO PAY WITH ANY BANK