近日，京东全新推出750B深度思考大模型、行业首批高商业可用数字人，并发布了角色智能体与具身 智能融合的最新进展，展示出AI认知能力与商业应用的最近进展。 "大模型的能力光谱在变、参数尺寸在变，不变的是让产业用好大模型。只有深耕产业，让大模型在产 业里跑起来，才是最有价值的事情。"京东集团探索研究院副院长、京东科技人工智能业务部总裁何晓 冬表示。 技术拓宽大模型能力边界，支持企业构建专有模型 在AI进化的"马拉松"中，京东大模型展现出了独特的产业基因。 目前，京东已经具备全尺寸大模型，满足多样化的产业需求：3B和10B模型可提供极致响应效率；81B 主力模型可兼顾效果和性能。最新推出的750B超大规模模型，则兼具"深度思考"和"非深度思考"双通 道能力，能满足各行业对"即时响应"和"深度推理"的双重需求。 具体来说，750B超大规模模型在训练过程中，使用了动态分层蒸馏、跨领域数据治理等京东创新技 术，降低大模型的训练和部署成本他，同时兼顾大模型效果，保证大模型能力"大而精"。 数字人实现高商业可用，618面向商家全免费 在大模型"硬实力"的支撑下，数字人等"软实力"应用，也迅速实现了商业化可用。 基于通用数字 ...

Core Insights - JD.com has launched a new 750B deep thinking model and the first commercially viable digital humans in the industry, showcasing advancements in AI cognitive capabilities and commercial applications [1] - The company emphasizes the importance of integrating large models into industries to maximize their value, as stated by He Xiaodong, Vice President of JD Group Exploration Research Institute [1] Group 1: Large Model Capabilities - JD.com has developed a full-size large model that meets diverse industry needs, with models of 3B, 10B, and 81B providing varying levels of response efficiency and performance [2] - The newly introduced 750B model features both "deep thinking" and "non-deep thinking" capabilities, catering to industries requiring instant responses and complex decision-making [2] - Innovative techniques such as dynamic hierarchical distillation and cross-domain data governance have been employed to reduce training and deployment costs while maintaining model effectiveness [2] Group 2: Technological Innovations - Recent research published in a Nature journal addresses the efficiency of large model development in open environments, introducing four core innovative methods: model distillation, data governance, training optimization, and cloud-edge collaboration [3] - These innovations have improved inference efficiency by an average of 30% and reduced training costs by 70%, creating a reusable industrial-grade technology paradigm [3] - The JD JoyBuild platform supports over 100 algorithms and toolchains, enabling businesses to quickly adapt general models into specialized ones [3] Group 3: Commercialization of Digital Humans - JD.com has achieved high commercial viability for digital humans, launching a general digital human model 2.0 that supports fine-tuning and natural expressions [4] - During the 618 shopping festival, JD.com offered six industry-specific digital humans for free to merchants, enhancing their sales capabilities [4] - A collaboration with a well-known snack brand resulted in a customized AI live streaming experience, generating over 10 million yuan in sales [4] Group 4: Embodied Intelligence - JD.com is exploring the integration of its large models into the physical world through the Joy Inside initiative, aiming to develop "personified" robots [5] - The Joy Inside platform utilizes over 10 million daily intelligent conversations to embed conversational AI into hardware like robots and AI toys, creating emotional connections with users [6] - This initiative is expanding the presence of intelligent interactions in everyday life, enhancing user experiences [6]

京东探索研究院 投稿 量子位 | 公众号 QbitAI 京东探索研究院关于大模型的最新研究，登上了Nature旗下期刊！ 该项研究 提出了一种在开放环境场景中训练、更新大模型，并与小模型协同部署的系统与方 法 。 它通过模型蒸馏、数据治理、训练优化与云边协同四大创新，这个项目 将大模型推理效率平 均提升30%，训练成本降低70% 。 这个名为《Omniforce：以人为中心的、赋能大模型的、云边协同的自动机器学习系统》的 项目，发表在Nature旗下期刊npj Artificial Intelligence上。 据介绍，这是国内首个系统性解决开放环境下大模型开发效率难题并获国际顶刊认证的研究 成果。 提出四个创新方法，推理平均提效30% 以京东大模型为例，蒸馏后的大模型Livebench提升14分。 大量的实验结果也证明有效性和效率， 推理平均提效30%，训练成本平均降低70% 。 根据企业自身业务，将通用模型转化为专业模型 企业将大模型应用付诸实践，面临着诸多卡点： 一方面进入大模型应用门槛高，另一方面模型训练与推理效率低。 京东大模型开发计算技术，能支持企业的模型开发训练及生产，让庞大、重型的AI模型"瘦 ...

随着大模型应用的深入，对企业的基础设施带来一系列全新要求和挑战。比如，以中央处理器（CPU）为中心的 架构在支持人工智能原生应用上面临挑战，需要以图形处理器（GPU）为中心重塑基础设施；面对激增的推理需 求，计算资源需求持续增加，企业需要思考资源投入产出问题。为此，面向大模型应用部署需求，京东云也提供 了多场景、多形态、多规格的解决方案。 曹鹏认为，随着大模型全面走向深度应用，企业当下需要做好三件事：分钟级部署+零门槛接入，大模型一体机 助力企业快速尝鲜大模型；深度应用全面开启，智能体重塑人工智能生产力；技术体系迎来全面重构，人工智能 基础设施走向标准化。 一体机作为最快速部署大模型的方式之一，是尝鲜企业级人工智能的最佳路径。过去三个月，"开箱即用"的京东 云大模型一体机快速发展，全国规模化落地已突破500台。京东云20日当天发布了三大垂直行业一体机，包括医疗 一体机、工业一体机、金融一体机。 在曹鹏看来，虽然大模型"超级应用"还有距离，但聚焦企业端的"深度应用"已奔涌而至，正在加速渗透到需要投 入大量人力进行重复劳动的场景。随着大模型及智能体技术的持续升级，行业正加速迈向深度应用阶段。 数据显示，近三个月京 ...

Core Insights - JD Cloud launched nine products including the JoyScale AI computing platform, JoyBuild large model development platform, and JoyAgent intelligent agent, aimed at helping enterprises reconstruct AI infrastructure and accelerate deep application adoption [1] - The company emphasized that the employment rate of digital employees will become a standard for measuring enterprise advancement, indicating that the extent of AI integration will determine future operational speed [1] - The new generation of agents, represented by JD Cloud's JoyAgent 2.0, is designed to assist enterprises in generating specialized digital employees, marking a significant step towards large-scale application and standardization of AI infrastructure [1]