Ant Lingbo Embodied Model
An embodied AI company under Ant Group focused on building cross-embodiment, cross-scenario robot brains, releasing the industry's first embodied-native world action model
In-depth Report
-
Ant Lingbo is a wholly-owned subsidiary of Ant Group established in December 2024. It specializes in the field of embodied intelligence and is positioned as a "universal brain supplier for robots" rather than an ontology manufacturer. In July 2026, Ant Lingbo continuously released and open-sourced six large embodied models within one week during WAIC, proposing the "Embodied Native" technical route—rejecting migration and fine-tuning on existing large models, but pre-training from scratch around the causal laws of the physical world. Its representative model, LingBot-VA 2.0, achieves single-card 150Hz real-time inference and has been put into actual operation in the Shanghai store of Guoda Pharmacy. Ant Group has formed a unique "light on body, heavy on brain" strategy on the embodied intelligence track through a three-tiered layout of self-research by subsidiaries, joint ventures (with Zhiyuan Robotics) and equity investments (covering 12 robotics companies).
-
The birth of Ant Lingbo began with a deep dismantling of the AI strategy within Ant Group. Ant Group divides its AGI goals into two levels: digital intelligence and physical intelligence. Digital intelligence is carried by Bailing Model, AI Alipay, Ant Afu and other businesses; the core executive body of physical intelligence is Ant Lingbo. In December 2024, Ant Lingbo was registered in Zhangjiang, Shanghai, with a registered capital of 100 million yuan and is wholly owned by Ant Technology Group Co., Ltd. CEO Zhu Xing joined Alibaba in 2011 and has more than 14 years of experience in payment, financial services, O2O and other fields. In 2023, he just completed the commercialization of Alipay advertising from 0 to 1. At the end of 2024, he was transferred to prepare for the establishment of Ant Lingbo. Chief Scientist Shen Yujun comes from the Interactive Intelligence Laboratory of Ant Technology Research Institute. He has an undergraduate degree from the Department of Electronic Engineering of Tsinghua University and a Ph.D. from the Department of Information Engineering of the Chinese University of Hong Kong. He has worked as an intern at SenseTime and as a senior researcher at ByteDance. At present, the Ant Lingbo team has about 200 people, and more than 90% of them have master's or doctorate degrees. Within Ant Group, Lingbo hangs under the CTO line, parallel to the Bailing large model team and Alipay AI team.
-
Ant Lingbo has built a complete technology stack from perception to execution. In July 2026, six models were released continuously within one week, covering six major directions: LingBot-Vision (visual basic model, 1B parameters): Using the boundary-centered self-supervised ViT architecture, RMSE leads Meta DINOv3 on NYU-Depth v2, surpassing the latter with only one-third of the data amount. LingBot-Depth 2.0 (spatial awareness model): Training data scaled from 3 million to 150 million samples, ranked first in 12 out of 16 depth completion benchmarks, RMSE dropped from 0.132 to 0.062. LingBot-VLA 2.0 (visual-language-action model, 6B parameters): The total amount of pre-training data reaches 60,000 hours (from 50,000 hours of real robot data + 10,000 hours of human operation data), and has been adapted to 20 robot configurations (single-arm, double-arm, bipedal, wheeled, etc.) from 17 manufacturers. In the GM-100 evaluation, the overall average task progress score and success rate are both ahead of π0.5 and GR00T N1.7. LingBot-World 2.0 / World-Infinity (real-time interactive world model, 14B/1.3B parameterized version): supports hour-level unlimited generation without visual drift, 720p/60fps real-time output, the industry’s first introduction of the Agent mechanism, and is all open source. LingBot-VA 2.0 (Embodied Native World Action Model): This is the flagship model released this time. It does not follow the industry mainstream route of "video generation model fine-tuning", but is pre-trained from scratch based on an autoregressive architecture. The core design includes: semantic visual-action tokenizer (incorporating semantic and action information alignment into visual compression), strict causal pre-training paradigm, MoE architecture (hybrid expert model), and enhanced asynchronous reasoning mechanism (executing actions while predicting future states). This model achieves a single-card 150Hz real-time inference efficiency, that is, 150 inferences per second, and a single response of approximately 6.7 milliseconds. LingBot-Video (video generation basic model): the world's first open source video generation basic model based on MoE architecture and oriented to embodied intelligence.In terms of practical application, Ant Lingbo has cooperated with Guoda Pharmacy to implement a robot smart pharmacy solution in Shanghai stores. Three robots of different configurations autonomously divide work driven by LingBot-VLA 2.0, completing the entire closed-loop process from order taking, picking up goods to delivery. This solution was selected as one of the top ten "treasures of the museum" in WAIC in 2026. In terms of hardware, Ant Lingbo launched its first service robot, Robbyant R1, in September 2025, targeting service scenarios such as home, elderly care, and medical health. It has been put into use in places such as the Shanghai History Museum. R1’s joint modules come from Titanium Tiger Robots, and its chassis comes from Galaxy General Motors, two Ant-invested companies. The R1 is still being tested primarily in controlled environments and has not yet been sold to retail customers.
-
Ant Lingbo’s core business model can be summarized as “selling brains, not bodies.” The company does not take the hardware manufacturing route, but builds a "public base" for embodied intelligence through open source and open model ecology. Specific pivots of its business model include: The first is to build an ecosystem through open source. All six models are open source (weights, codes, post-training tool chains, and evaluation sets are all released), attracting global developers and robot manufacturers to access through open source, forming a developer ecosystem with the Lingbo model as the core. At present, more than a dozen robot manufacturers have established cooperation with Lingbo. The second is scene-based charging. Although specific pricing has not been announced, the business model for implementation scenarios such as pharmacies, retail sorting, and logistics sorting has begun to be verified. Ant Lingbo cooperates with ecological customer partners such as Guoda Pharmacy and Longsheng to provide "brain" services without renovating stores. The third is the joint construction of data ecology. Ant Lingbo and Jianzhi Robot have reached a strategic cooperation to jointly develop exclusive data mining equipment. As a core data partner, Leju Robot has provided nearly 10,000 hours of high-quality multi-modal real robot data for LingBot-VLA. Ant Group's overall layout does not stop at Lingbo itself: through its subsidiary self-research (Robbyant), joint ventures ("Hangzhou Chuanzhi Future Technology Co., Ltd." jointly established with Zhiyuan Robot, registered capital of 20 million yuan), and investments in 12 robot companies within 18 months (the latest one led Zhiyuan Robot's 500 million yuan financing round), it has built a multi-layered and risk-dispersed strategic layout.
-
Ant Lingbo’s technical route has caused obvious divisions among developers and industry practitioners. A partner manufacturer said, "Lingbo's model gives us a new choice without having to build a brain from scratch." Leju Robot officially announced that its KUAVO 4 Pro (Kuafu) has completed the post-training adaptation of LingBot-VLA and conducted a systematic evaluation in 95 real operation scenarios. But some practitioners in the robotics industry have raised questions. "The inference frequency is only one of the engineering indicators. The real test is the generalization ability in real scenarios. Running 150Hz in a simulation environment and running 150Hz in a home environment with elderly people, children, and pets are two different things." Regarding the R1 robot, user feedback is still in the early stages. It is currently mainly used in museum tours, pharmacies and other scenarios, and there are no large-scale consumer experience reports yet. Some viewers who visited WAIC said that Lingbo's table tennis demonstration was "impressive, and the robot's reaction speed is really fast." However, some people thought that "there is a huge gap between the demonstration environment and real home scenes."
-
Ant Lingbo has received high attention from industry media and capital markets. Geek Park’s evaluation: “Ant Lingbo leads the overall performance in the GM-100 evaluation, showing stronger dual-arm collaborative operation capabilities and cross-body and multi-task generalization capabilities.” 36Kr’s in-depth analysis believes that “embodied native represents a technical path that is different from the mainstream of the industry, and cuts into an underlying issue that has been ignored for a long time – whether robots should have their own model design paradigm.” AI Dev Signals ranked Ant Lingbo second in the world in the field of Physical AI (after NVIDIA) and evaluated it as "the most radical open source promoter in this field." In terms of competitive product landscape, Ant Lingbo’s “light on body, heavy on brain” route is in sharp contrast to Figure and Tesla’s vertical integration route. Goldman Sachs is cautious about the commercialization process of humanoid robots in its research report. It believes that the technological inflection point is still unclear. It is expected that global humanoid robot shipments will be about 76,000 units by 2027 and about 502,000 units by 2032. The pace is slower than market expectations.
-
The challenges facing Ant Lingbo are equally clear. The biggest uncertainty comes from the technical route itself. The path of "embodied native" pre-training from scratch means higher R&D investment and longer verification cycle. Compared with the "results are produced in a few months" rhythm of the migration fine-tuning route, Lingbo's route will take longer to prove the advantage of its generalization ability - if the final effect does not significantly exceed the migration plan, the high investment will be unsustainable. Data bottlenecks are a more realistic difficulty. The embodied intelligence industry faces a fundamental problem: there is a gap of tens to millions of times between real-world data and language data. Lingbo's data source mainly relies on the supply of ecological partners, and it does not create a large number of ontologies itself. But the problem is - Yushu Technology has just launched its own large embodied model UnifoLM-OminiA-0.3, and Zhiyuan Robot also has its own world model and data platform. Once these ontology manufacturers start seriously developing their own brains, the data supply may be tightened at any time. Although open source is a strategy for building an ecosystem, it does not itself constitute a moat. Within Ant Group, Lingbo needs to compete for resources with the Bailing Model Team and the Alipay AI Team. Outside Ant Group, within the Alibaba department alone there are multiple embodied intelligence lines that are "fighting independently," including Qianwen Qwen-Robot and AutoNavi Model, but have yet to form a unified strategic synergy. The shipment volume of the R1 robot is also relatively limited. It is currently closer to a "carrier for verifying the model" and is still far away from large-scale mass production.
-
Ant Lingbo's products are currently mainly targeted at two types of audiences: First, robot body manufacturers can reduce repeated investment at the algorithm level by accessing Lingbo's universal brain, which is especially suitable for small and medium-sized robot companies. Second, operators of specific service scenarios, such as pharmacies, elderly care institutions, museums, etc., can quickly deploy robots with autonomous decision-making capabilities through Ant Lingbo’s solution. For ordinary consumers, Ant Lingbo's products are still in the early stages. The purchase channels and official pricing of the R1 robot have not yet been made public. It is not recommended for individual consumers to buy it for the time being. If you are looking for alternatives, Yushu Technology’s humanoid robots G1 (priced at approximately US$16,000) and H1 (approximately US$90,000) have obvious advantages in terms of hardware cost performance, but require self-built algorithm capabilities at the “brain” level.
-
Ant Lingbo represents an important step in Ant Group’s AI strategy—from the digital world to the physical world. Its "embodied native" technology route is a bold attempt. If successful, it may redefine the development paradigm of robot brains; if verification is not smooth, high R&D investment and the "moatless" dilemma of the open source ecosystem will become hidden worries. In any case, Ant Lingbo's style of play is unique among major domestic manufacturers. It answers not "how to make a better robot", but the lower-level question "where should the robot's intelligence come from?"
User Reviews
-
CelestiaSky360—Saw Ant LingBo's smart pharmacy demo live at WAIC — three different robots collaborating to fetch medicine, no human intervention needed, and the on-site effect was genuinely stunning. Guoda Pharmacy is already using it; it's not a lab product, and that matters. -
Sarah.Murray5209—LingBo's LingBot-VLA 2.0 is open-sourced — pretrained on 60,000 hours of real-robot data, supporting over 20 robot configurations. That data volume has no rival in the open-source world right now; it even surpasses Physical Intelligence's π0.5. -
BenjaminGonzalez_7—Honestly, LingBo's open-source strategy is pretty smart. The tech route isn't set in stone yet, so whoever builds the ecosystem first gains thengôn quyền. Just not sure how the monetization path goes — selling models and selling hardware are two different things. -
ABrown007—150Hz real-time inference running on a single card — the efficiency is genuinely competitive. But reading the tech docs, how different their 'embodied-native' really is from the fine-tuning route still depends on actual deployment results. -
Betty.MoralesZ663—I've been following Ant's LingBo for a while. CEO Zhu Xing came from Alipay's commercialization side, with no robotics background, yet chose the hardest path of pretraining from scratch. This team's technical judgment is genuinely worth watching. -
AJohnson_8893—Running at 150Hz in a simulation environment is one thing, but it's another at home, where there are elderly people, kids and pets, and enormous environmental uncertainty. LingBo's pharmacy scenario works for now, but it's still far from entering the home. -
GP_wal—Saw the table-tennis rally demo at WAIC — the robot's reaction speed is genuinely fast; the moment the ball bounced it was already predicting where it would land. No external high-speed camera, purely onboard vision. That's pretty impressive. -
SCwar—Ant's LingBo route is completely different from Unitree's: one builds the brain, the other the body — they're not really competing. The problem is LingBo's data depends on body makers; the day Unitree or AgiBot build their own brain, the data source dries up. -
Anthony.Howard—Six models open-sourced within a week — that kind of boldness is rare. From the deep learning framework to the training toolchain, everything's out there, and the developer community's response has been good. But open-sourcing itself isn't a moat; iteration speed is. -
SHill_Plus68—I've seen the R1 robot giving tours at the Shanghai History Museum — its walking was fairly steady and the interaction was okay. But a robot backed by 100 million in registered capital, next to Tesla or Figure with billions in funding, is just on a different scale. -
CherylKim_2021—A friend who builds robots told me they're adapting to LingBo's LingBot-VLA. Where they used to need a dedicated algorithm team for their robot, now it just plugs into LingBo's 'brain' and moves. For small companies, that saves a fortune. -
JBailey_Plus7—Essentially it's Ant Group's technology reserve for offline scenarios. Payments, healthcare and credit are already deeply embedded offline, and robots are the natural extension. Not a pivot, but an extension of the core business. -
BreadBud513—LingBot-Depth 2.0 improved a lot on transparent object recognition — things like champagne flutes, impossible for depth cameras before, now get a complete outline reconstructed. That's critical for grasping tasks. -
赵华_1—A lot of people hype embodied AI, but after reading LingBo's materials I realized this track is far harder than imagined. Real-world data is millions of times larger than language data; just solving the data problem alone will burn years of money. -
SusanEdwards—Over 90% of LingBo's team hold master's or PhD degrees. Chief scientist Shen Yujun did his undergrad at Tsinghua and PhD at CUHK, with a computer vision background. The technical foundation is solid, but a 200-person team is still tiny compared to Google DeepMind. -
琉璃_23—Read the Goldman Sachs report — they're pretty cautious on robot commercialization, projecting only 76k units of global shipments by 2027. Ant is pouring money into LingBo with no short-term returns in sight, but long term, if you don't position early you get left behind. -
EHall520—Turns out LingBot-Video uses an MoE architecture — 30B parameters but only 3B activated at inference. Pretty clever design. It's a video model built specifically for embodied AI training, a completely different route from content generators like Sora. -
Sandra156—Just watched LingBo's launch event. VA 2.0 was pretrained from scratch instead of fine-tuning an existing model — that approach really takes guts. But the results speak for themselves: the demo of picking up a potato chip without crushing it was incredibly convincing. -
Brenda.Ross_Plus—Ant has invested in a dozen-plus companies on the robotics track — Unitree, Galaxea and Lingxin Qiaoshou all have its stakes. LingBo itself doesn't build robot bodies; it positions as a brain supplier, which is genuinely one of a kind among big Chinese players. -
dm0feb8—Walked through the embodied AI zone at WAIC and LingBo's booth had the biggest crowd by far. The 'deploy without renovating the store' pitch really lands — lots of pharmacies and supermarkets want robots but don't want to remodel their space.