Ant Lingbo's LingBot Appears at Wuzhen: LeRobot, vLLM and timm Support, but Verify the Scope of Openness First
Leiphone reported on November 3, 2026 that Ant Lingbo showcased the LingBot series of robot brains at the World Internet Conference in Wuzhen, saying the models had been adapted to LeRobot, merged into the timm mainline and gained vLLM and SGLang support. The report did not disclose licenses, model sizes or benchmark data.

What changed
The World Internet Conference Wuzhen Summit was held from November 2 to 5, 2026. Ant Group's embodied intelligence company Ant Lingbo exhibited what it calls a 'full-stack open source' general-purpose robot brain at the event. According to Leiphone's November 3 report, the series is called LingBot; the specific models named in the report include LingBot-VLA, LingBot-VA, LingBot-Vision, LingBot-Map and LingBot-World, with an image caption reading 'Ant Lingbo Brain Full-Stack 2.0 Model Matrix'. The report said the LingBot series had been open source for nine months, had surpassed 2.2 million downloads and 33,000 GitHub stars, and that related work had been cited by teams including Physical Intelligence and Generalist AI. The most concrete part of the report concerns toolchain integration: LingBot-VLA and LingBot-VA were said to have completed adaptation to the robot development platform LeRobot; LingBot-Vision was said to have been merged into the mainline of timm, a mainstream vision backbone library; and LingBot-World was described as the world's first world model supported by both of the open-source inference engines vLLM and SGLang, and as having been integrated by Nvidia into its open-source world-model inference library FlashDreams. The report also listed several external uses: Stanford University's OpenWAM project reused and adapted the LingBot-VA module; a South Korean FastOPD study used LingBot-VLA as a teacher model; and EPFL in Switzerland used LingBot-Map for 3D reconstruction of coral reefs. These are all as stated in the report, with no independent reproduction described. On the application side, Ant Lingbo and Guoda Pharmacy jointly demonstrated a smart pharmacy in Wuzhen: the booth recreated a real pharmacy environment, and a robot completed a continuous operation from order receipt and intelligent sorting to product delivery in an 80-centimeter-wide aisle; the report said the solution has been operating routinely in multiple Shanghai Guoda Pharmacy stores and requires no changes to product display or store layout. The report also mentioned expanding industrial loading and unloading and logistics sorting with partners including Longsheng Weirui, Dema Technology and De'er Xiangwang; validating 'one brain, multiple robots' with robot-body manufacturers including Leju and Taihu; and signing a memorandum of cooperation with the Arab Digital Economy Alliance. The online preliminary round of the first Ant Lingbo Embodied Large Model Challenge was said to have attracted 1,855 teams from 648 universities and companies in 13 countries; the full-stack general-purpose robot brain models and the smart pharmacy received the conference's 'Xinguang' Product Award and 'Xinyao' Scenario Award, respectively.
What it means for developers
For developers, the substantive change is not in parameters but in the integration point. Embodied models have often appeared as paper code or single-point demos, and teams had to assemble data formats, action spaces and inference backends themselves. The LeRobot adaptation, timm mainline merge, and vLLM and SGLang support described in the report amount to placing LingBot into three existing engineering habits—the training and evaluation loop, the vision backbone registry, and the model-serving layer. If these integrations are actually usable, the starting point for evaluation shifts from 'rewriting a pipeline' to 'getting an existing pipeline running and then swapping in the model'; what is usually most expensive in embodied projects is data collection and real-robot debugging. But 'integration completed' and 'good enough once connected' are two different things. The report gives no version numbers for the integrations, supported observation and action spaces, control-frequency assumptions, VRAM use or throughput figures, and it does not say at which layer the adaptation happens: data format, policy wrapper, or replacement point in the training script. Another point: world-model inference workloads differ greatly from language models; sequences are longer and more memory-sensitive, so being able to start a service does not mean meeting the latency budget of a control loop. The 80-centimeter narrow aisle, dense placement and continuous operation in the smart pharmacy are useful scene constraints, but the number of stores, continuous runtime, manual takeover rate and failure cases are not disclosed, so they cannot support inferences about success rates in other retail or warehouse environments. 'One brain, multiple robots' is an ongoing validation with robot-body makers such as Leju and Taihu, not a reproducible cross-embodiment transfer result.
Integration and evaluation
These are conditional recommendations and analysis, not reported facts. Start with one upfront check: the report repeatedly calls the series 'full-stack open source', but it names no license and does not say whether what is open is training code, inference code, weights, or all three. Before investing evaluation resources, confirm the scope of openness and commercial terms with the official repository; download and star counts are no substitute for a license grant. First, if you already use LeRobot for training and evaluation, and the official project does provide accessible weights or code, you can choose to connect LingBot-VLA or LingBot-VA to your existing LeRobot evaluation process as a complete model or policy, rather than assuming it can replace a particular backbone network. The precondition is to check the official documentation to confirm whether the adaptation level, observation dimensions, control frequency and action representation match your robot body, and whether additional wrapper code is needed. Evaluation criteria can be: under a fixed fine-tuning step budget, whether success rate on your own tasks rises or falls relative to the current policy; the time from obtaining a usable model to the first successful rollout; and whether per-step inference latency falls within your control-cycle budget. If the first rollout is clearly faster but the success rate drops, the saving is engineering time, not model capability. Second, if your serving stack already has either vLLM or SGLang, and the workload involves world-model-style long-sequence generation, you can choose to start with the engine you are familiar with, but you must first verify that the version you use actually includes support for LingBot-World—the report only says 'supported' and gives no starting version. The Nvidia FlashDreams integration is likewise a report claim and needs confirmation in the corresponding repository. Evaluation criteria should be end-to-end latency at a fixed batch size, the number of concurrent requests per GPU and VRAM headroom, with an exit criterion: if the target sequence length exceeds VRAM or latency budgets, fall back to offline generation rather than online serving. Third, if your perception side is already built around timm, you can attach LingBot-Vision as a frozen backbone to your existing task head for linear probing or lightweight fine-tuning. The precondition is that input resolution and the preprocessing pipeline can be aligned, so you do not change the whole data pipeline for one backbone. Evaluation criteria are downstream accuracy and per-frame latency on a small batch of real-scene frames you have labeled yourselves; replace the existing backbone only if the gain is enough to cover switching and regression-testing costs. All three rules share the same point: first get a small-scale run working on your own data and robot body, then decide whether to expand investment.
Limits and open questions
First, distinguish 'open source' from 'open weights': the report uses 'full-stack open source' to describe the LingBot series, but it names no license anywhere and does not say whether the release is code, weights or both, so it cannot be used to infer rights to commercial use, modification or redistribution. Second, the report gives no model sizes, parameter counts, observation windows, action-space definitions, training-data composition or benchmark scores, and there is no third-party reproduction; the citations and integrations in the article (including citations by peer companies, reuse in university projects, and Nvidia's FlashDreams integration) are statements by the releasing party or the report, not the same as independent validation of capability. Third, the deployment descriptions for the smart pharmacy and 'one brain, multiple robots' come from the company's own account and lack verifiable metrics such as store count, continuous runtime, manual takeover rate and failure cases. Fourth, the report does not mention a hosted API, a commercial licensing path or hardware requirements, so these need confirmation with the releasing party before actual deployment. Finally, it should be clear: LingBot-VLA, LingBot-VA, LingBot-Vision, LingBot-Map and LingBot-World are not currently in KimiSeek's model catalog, and KimiSeek has not tested them; the downloads, integrations and store deployments mentioned in the article all point to Ant Lingbo and its partners, and any availability claim should be checked against their official channels—do not assume they can be called directly on this platform.
Illustrations from the original report
