JD Yanxi 3

• 1. overview (712 chars)

  JD Yanxi 3 is the third generation form of JD Group’s self-developed industrial AI platform, rooted in the Yanxi large model, Digital Human 3.0 and the upgraded JoyAI brand matrix in 2025. It transforms JD.com’s supply chain, customer service and e-commerce data accumulated over more than ten years into multi-modal capabilities such as language, voice, image, video, and digital people, focusing on serious business scenarios such as retail, finance, government affairs, and logistics.   Different from general consumer-oriented large models such as Tongyi, Wenxin, and Doubao, Yanxi's differentiation lies in its "following the industry" - training with 70% general data and 30% JD.com digital supply chain data, and outstanding performance in high-knowledge-density tasks such as e-commerce, customer service, and live streaming.

• 2. Product background (1501 chars)

  The predecessor of Yanxi was the manual customer service team established by JD.com in 2009. It officially launched an intelligent customer service brand in 2020 and upgraded to a 2.0 platform in 2022. In July 2023, JD.com launched the Yanxi large model with hundreds of billions of parameters at the Explorer Conference and announced a "three-step" industrial implementation plan. In 2024, JD Cloud will release the Yanxi intelligent agent platform and the Yanxi digital human 3.0 platform. In February 2025, all Yanxi products were connected to DeepSeek R1/V3; in July of the same year, at the World Artificial Intelligence Conference, the Yanxi large model brand was upgraded to JoyAI, forming a full-size model matrix from 3B to 750B.   Entering 2026, JD.com has significantly accelerated the pace of AI commercialization. In January, JD.com established the "Chameleon" business unit to integrate AI projects that were originally scattered across various businesses into independent business units, upgraded the consumer-side large-model assistant JoyAI App, and launched Jingxi Shopping Assistant. In March, JD Cloud released the "Yanxi" government affairs big data model 3.0, which strengthens the correlation analysis of massive heterogeneous data in urban governance and supports one-click voice generation of social governance analysis charts. It has been implemented in smart city projects such as Beijing and Suzhou. On the 24th of the same month, JD.com released the Instruct version of JoyAI-LLM Flash, a large open source basic model, focusing on "Token efficiency". The Tokens consumed at the same task accuracy are only one-quarter to one-fifth of similar models, and are specially designed to cooperate with intelligent applications such as OpenClaw.

• 3. Product functions and usage experience (2377 chars)

  The core of Yanxi 3 is a set of large multi-modal models. The large language model supports model self-evolution and low-bit high-precision quantitative reasoning, which can save 70% of video memory; the large speech model is based on 200,000 hours of speech data training, supports zero-sample and cross-language speech synthesis, and can recognize Chinese, English, Thai, Indonesian, Vietnamese and other foreign languages ​​and 20 Chinese dialects; the large image model provides Chinese native textographic capabilities and supports ultra-high resolutions above 4K and customized styles; the large video model uses a self-developed 4D parallel framework, which can generate up to 32 Seconds of 1080p high-definition video; the digital NPC model features "zero illusion, zero-sample real-time reasoning" and can respond in real time in the live broadcast room.   The most mature application layer is digital human live broadcast. The Yanxi Digital Human 3.0 platform released in 2024 supports one-click broadcasting in the cloud without local computer configuration. Thousands of live broadcast rooms can be online at the same time, and the cost of broadcasting is reduced by 30% compared with before. By the first quarter of 2026, JD Digital People have served more than 70,000 merchants, and the number of broadcasts has increased 10 times year-on-year, driving tens of billions of GMV. At the JD Merchant Conference in April 2026, JD further announced an “interactive Agent" is the core of the digital human live broadcast intelligent interaction system, which introduces e-commerce knowledge to enhance question and answer, field control dynamic response and multi-agent collaboration, and proposes a "fast and slow dual-drive" evolution direction: fast thinking is responsible for millisecond-level basic interactions, and slow thinking deeply integrates supply chain, inventory, price and other full-link data to make business decisions, allowing digital people to move from "puppets listening to instructions" to retail agents that can independently arrange products, automatically replenish goods, and intelligently promote orders.   On the other hand, JD.com will launch an independent app "JD.com AI Shopping" at the end of 2025. The base is a Yanxi large model and focuses on conversational shopping. In May 2026, a number of media conducted actual tests and found that it "can be used in product recommendation, AI try-on, and take-out recommendations, but it cannot replace people": the try-on pictures cannot restore the true body shape, take-out recommendations often do not understand "nearby", and the regular repurchase function is not stable enough. However, in terms of price comparison and low-price monitoring, JD.com AI Shopping performed better than its competitors.

• 4. Pricing and business model (704 chars)

  Digital people are Yan Xi’s smoothest monetization path at present. In December 2025, JD Digital People will be fully free and open. The software system is free and the average daily cost is less than 100 yuan. Small and medium-sized businesses can start broadcasting with almost no threshold. On the enterprise side, there are also JoyAI enterprise agents (JoyAgent), commercial licenses, and the Yanxi AI development computing platform that is billed based on training and inference.   For the implementation of intelligent agents, JD Cloud has launched the OpenClaw all-in-one machine based on domestic chips, which can be used out of the box and comes with a Token package to reduce computing power costs. Overall, Yanxi’s commercialization focuses on “free tool traffic, enterprise services and computing power fee realization”.

• 5. User reviews (1056 chars)

  Feedback from real users and developers on Yanxi is quite divided. In e-commerce and customer service scenarios, Yanxi is generally considered "down-to-earth": someone used JD Yanxi and DeepSeek each for a week and concluded that Yanxi is more proficient in Chinese e-commerce dirty strings (SKUs, orders, comments). The permissions, billing, monitoring, and privatization deployment of JD Cloud Family Bucket are all ready-made, and the enterprise compliance link is also smooth. The cost reduction effect of digital human live broadcast is the most recognized - the cost is only one-tenth of that of a real team, and the conversion rate can exceed 30% during leisure time.   But the shortcomings are equally obvious. Non-JD Cloud users have greater friction in accessing the Yanxi ecosystem; general code tasks are not top-notch, and the gap with DeepSeek is mainly outside the e-commerce scenario; the documentation is more scattered than Tongyi Qianwen, and there are not enough community cases; some advanced functions need to be applied for separately; pure individual users will feel that the price is not cheap enough. The AI ​​shopping assistant has been named by many media as "stability, trust, and user habits have not yet been established."

• 6. Industry evaluation (791 chars)

  In the pattern of large industrial models, Yanxichang is juxtaposed with Tongyi Qianwen, Wenxinyiyan, and Huawei Pangu. Tongyi is leading in multi-modal generation, Pangu Industrial has high quality inspection accuracy, and Wenxin government resources are excellent in integration. Liexi’s strength lies in supply chain optimization. At JD.com’s 2025 financial report, management revealed that the number of calls to the self-developed Yanxi large model has surged nearly a hundred times in one year, and the number of JD Cloud Token calls has hit a record of weekly growth of 455%.   However, some analysts believe that the overall AI layout of JD.com is "laggard and conservative": when Alibaba fully integrated Qianwen into Taobao and Byte Doubao's daily sales exceeded 10 million, JD.com did not establish an independent AI business department until January 2026, and its investment in C-end products was obviously insufficient.

• 7. Controversies and Risks (681 chars)

  The biggest controversy is "whether digital people will replace real anchors." JD.com's own positioning is very restrained - digital people are a complement to the real-person team, and what they do is to tap the long-tail traffic after the real-person broadcast. The average idle conversion rate exceeds 30%, rather than direct competition.   Another risk comes from the actual performance of AI shopping: inaccurate recommendations, inconsistent prices and actual transaction prices, strict platform channel barriers, and irrelevant products are often mixed in. These shortcomings make Yanxi still far away from being a “caring housekeeper” on the C-side. In addition, JD Cloud products are deeply tied to its own ecosystem, which is not friendly enough for companies that want to mix multiple models.

• 8. Applicable groups and usage suggestions (793 chars)

  Yanxi 3 is most suitable for B-end customers with serious business scenarios: retail brands provide digital live broadcasts and intelligent customer service, financial institutions provide digital employees for risk control, government departments provide 12345 hotlines and urban governance, and logistics and manufacturing companies provide process automation. For small and medium-sized businesses on JD.com’s platform, digital live streaming is the most cost-effective entry point at the moment, as it is free and can make up for spare time.   Ordinary consumers can try JD.com’s AI shopping for price comparison and repurchase, but don’t expect it to completely replace manual selection. If you are pursuing general code capabilities or the lowest price, open source models such as DeepSeek will be more cost-effective; if you are deeply involved in supply chain and e-commerce scenarios, Yanxi is still the more right choice.

• 9. Summarize (448 chars)

  JD.com Yanxi 3 is not another general chat model chasing hot topics, but an industrial AI platform that uses JD.com’s supply chain and transaction scenarios as a moat. It has already passed through the closed business loop on the B-side (digital human live streaming, enterprise intelligence), and is still in the cultivation stage on the C-side (AI shopping, JoyAI App, JoyInside hardware brain). For companies that want to use AI to really "work" rather than "chat", Yanxi is one of the most industry-friendly options at the moment.

No in-depth report yet

• 

  毛雪素

  —

  京东数字人真的省，我们小店开播成本一天不到一百块。

---

• 

  VWalker_77

  —

  我拿京东言犀和 DeepSeek 各用了一周做对比，结论是各有所长、谈不上谁吊打谁。言犀对中文电商业务细节明显更敏感，SKU、订单、评论这些脏字符串它吃过见过；DeepSeek 在通用代码和复杂推理上更稳，价格也更低、还开源。落地选哪个，最终还是看你的场景是不是贴着电商走。

---

• 

  韩丽

  —

  言犀文档比通义散，踩坑只能自己翻。

---

• 

  realSuzannaLucas_x

  —

  JoyInside 接进玩具里，我妈现在不催我电话了。

---

• 

  海浪_11

  —

  我们公司是京东云老用户，上 JoyAgent 做售后工单处理，原本要人工分类回复的内容现在智能体自动识别、自动派单，效率提升非常明显。但说实话，如果不是京东云体系内的企业，光是权限、计费的接入摩擦就够喝一壶，言犀这套生态绑得确实紧，好处是合规链路现成，坏处是迁移成本高。

---

• 

  史妍晴

  —

  做电商客服的，言犀对订单、评论这些中文脏字符串确实挑得动，比通用模型接地气，工单分类顺手多了。

---

• 

  JesusScott

  —

  京东 AI 布局晚了半拍吧。

---

• 

  Billy_Reed

  —

  我们做家电的，之前请真人主播一场下来成本好几万，还得排班。换成言犀数字人后云端一键开播，上千个直播间同时跑，单店成本降到两位数，闲时转化反而比真人高。现在数字人已经是我们标配了，但复杂场景的互动还得运营盯着，离「完全自主经营」还差口气。

---

• 

  Jennifer_LeeK

  —

  我们品牌用言犀数字人做了「直播间复刻」，上传一段真人直播视频就生成了形象声线高度一致的数字分身，不用再请主播熬夜播。上线后公域流量提升超六成，人均停留时长接近两分钟，对人力有限的中小商家而言，这套工具的性价比确实没话说。

---

• 

  KathleenKim_2022

  —

  用言犀做智能外呼，私域加粉效率高了不少。

---

• 

  EugeneBaileyK913

  —

  实事求是讲，京东 AI 购现在还是「能用但别指望替代人」。我让它推荐减脂期附近的轻食外卖，它推的品类单一，门店还都在五公里开外，明显没吃透「附近」这个需求；定期复购功能我也测了，到点经常不触发。整体体验离贴心管家还有距离。

---

• 

  PaulBrownZ

  —

  京东 AI 购试穿图根本看不出我梨形身材，参考价值不大。

---

• 

  Cebai

  —

  数字人直播适合标品和闲时增量，凌晨那种没人播的时段反而能出单，我们农产品店凌晨 GMV 占比直接从 3% 拉到 27%。

---

• 

  JerryKruger

  —

  东哥数字人首秀那场太魔幻了。

---

• 

  Joseph_Smith_Max

  —

  安踏那场数字人直播间 GMV 破两千万、转化率三十多，基本每三个进直播间就有一个下单，这数据确实能打。

---

• 

  Kevin.Torres

  —

  政务大模型 3.0 那个语音生成社会治理图表的功能挺实用，我们街道试点后报表时间省了一大半，比人工整理强太多。

---

• 

  NFlores_X

  —

  看过京东 JoyAvatar 的论文，长时长、自由态数字人这次确实突破了「站桩播报」，复杂指令能丝滑做全身动作，唇形还能和音频完美同步，GSB 评分超过 omnihuman 和 KlingAvatar，技术上是代差级优势，做直播短视频的想象空间一下就打开了。

---

• 

  DWard_6632

  —

  价格不算最便宜，个人开发者用言犀会有点心疼，开源的 Flash 版才更香。

---

• 

  Kenneth.Gonzalez_356

  —

  言犀这波开源挺良心，Flash 模型消耗的 token 才别家的五分之一。

No reviews yet