LingBot-VLA 2.0 Review: The Free Robot Brain Behind 2.2M Downloads

Most robot brains are built for one body. A model trained on a single-arm rig in a lab usually falls apart the moment you bolt it onto a different arm, let alone a humanoid. That is the problem Ant Group's robotics arm, Robbyant, set out to solve with LingBot-VLA 2.0. At this week's World Internet Conference Wuzhen Summit, Ant Group said the LingBot series has now passed 2.2 million downloads since it was open-sourced about nine months ago. The model is free, the code is open, and the weights are downloadable today. What you get for that price is a capable robot policy, plus a set of benchmark numbers that deserve a closer read before you trust them.

What LingBot-VLA 2.0 actually does

LingBot-VLA 2.0 is a vision-language-action model, or VLA. That is a clunky name for a simple idea: the model watches the world through a camera, reads an instruction in plain language, and outputs the movements a robot should make next. Tell it to sort the cans in the fridge and it works out the reaching, grasping, and placing itself.

The model has 6 billion parameters and is built on Qwen3-VL-4B-Instruct, Alibaba's open vision-language model, with two teacher models distilled in during training: LingBot-Depth for spatial geometry and DINO-Video for how scenes change over time. Robbyant trained it on roughly 60,000 hours of data, made up of 50,000 hours of real robot trajectories spanning 20 robot configurations and 10,000 hours of first-person human manipulation video. The data covers machines from 17 manufacturers, including Unitree, AgileX, Galaxea, Franka, and Astribot, in single-arm, dual-arm, bipedal, and wheeled forms.

The headline feature is the unified action space. Instead of training a separate model per robot, LingBot-VLA 2.0 maps every machine into one 55-dimensional action vector covering arm joints, end-effector poses, grippers, dexterous hands, waist, head, and the mobile base. Train once, and the same policy can in principle drive any of the 20 configurations it learned. That is the whole pitch: one brain, many bodies.

What it costs: the model is free, the hardware is not

LingBot-VLA 2.0 costs $0. The code and weights are released under the Apache 2.0 license, which permits commercial use, modification, and redistribution without a licensing negotiation. There is no API to pay for, no subscription tier, and no per-task charge. You download the 6B checkpoint from Hugging Face or ModelScope and run it on your own machine, with no data flowing back to Ant Group.

The real bill is hardware. Robbyant's own deployment guide measures one inference call at about 130 milliseconds on an NVIDIA GeForce RTX 4090D, using 10 denoising steps, which is fast enough for real-time control. But that is a high-end consumer GPU, and the 4090D is a China-market card. Post-training your own tasks needs more: the worked example fine-tunes on 50 RoboTwin 2.0 simulation tasks, and the training notes reference single-node H20 GPUs. Then there is the robot itself, which dwarfs the GPU in cost. Local AI hardware is not cheap in general. NVIDIA's DGX Spark desktop starts at $4,999, for example. The honest way to read the price is that the software is free and everything around it is your budget.

The benchmark numbers, read honestly

Every performance number below comes from Robbyant's own repository and technical report. No independent lab has published a replication yet, so treat these as the vendor's best case, measured on its own tasks.

On GM-100, a bimanual manipulation benchmark from Shanghai Jiao Tong University, LingBot-VLA 2.0 beats the two best-known open rivals on both tested platforms. The figures are progress score / success rate, where success rate is the share of tasks completed end to end.

The lead over π0.5 is real but modest, about 7 points of progress score on each platform. The more important number is the success rate itself: 34.4% on the best platform means the model still fails roughly two out of three full tasks. That is normal for generalist robot policies in 2026.

On long-horizon mobile manipulation, multi-step jobs like sorting a fridge or cleaning a stove, the model holds up better than π0.5 in-domain, but the gap between familiar and unfamiliar setups is steep:

Move the same fridge task to an unfamiliar layout and the success rate falls from 60% to 13.3%. That collapse is the central fact of robot AI right now: these models work in the room they trained in and degrade fast outside it. Robbyant deserves credit for publishing the out-of-domain numbers at all, but they say this is a research-grade policy, not a product you can drop into a new warehouse and trust.

In simulation, the results look stronger. After post-training on RoboTwin 2.0's 50 tasks, LingBot-VLA 2.0 scores 93.52% on clean tasks and 92.80% on randomized ones, against 82.74% and 76.76% for π0.5. Simulation is the friendliest possible test, though, with no worn grippers or bad lighting. Treat it as evidence the training pipeline works, not as a promise about your robot.

How it stacks up against π0.5 and GR00T N1.7

The open robot-brain field has three serious options, and LingBot-VLA 2.0 is the newest of them. Physical Intelligence's π0.5 is the established open-weight baseline, also Apache 2.0, with a large community around its openpi repository. NVIDIA's GR00T N1.7 is a 3B-parameter open model aimed squarely at humanoids, under NVIDIA's Open Model License. Google's Gemini Robotics exists too, but it is closed, so you cannot download it.

On Robbyant's own benchmarks, LingBot-VLA 2.0 leads both rivals. But a November 2 study called RoboColosseum, which ran 5,699 blinded pairwise comparisons across 124 real-world sites, found that "all systems struggle to complete open-world tasks reliably" and that rankings change completely after fine-tuning. LingBot was not in that study, so the finding neither disproves Robbyant's numbers nor supports them. LingBot-VLA 2.0 looks like the strongest open VLA on paper, in a field where paper leads have not yet translated into reliable real-world performance.

How to try it yourself

The repository ships everything needed to run the model: install scripts, a download helper, post-training examples, and a deployment script. The documented path is short. Clone the repo, create the Python 3.12 environment with the provided script, then pull the weights:

python3 scripts/download_hf_model.py --repo_id robbyant/lingbot-vla-v2-6b --local_dir lingbot-vla

For live control, the deployment script serves the policy over a port and runs it against a real robot, with the 130-millisecond figure measured on an RTX 4090D. To teach it a new task, you prepare a LeRobot v2.1 or v3.0 dataset, define the robot's feature mapping in a YAML config, and run the provided post-training script. The worked example fine-tunes on RoboTwin 2.0's 50 tasks.

That is all doable for a robotics team with GPU access, but it is not a weekend project for a general user: you need a compatible robot, a Linux machine with PyTorch 2.8, and patience to debug the gap between a simulation demo and a physical arm. If you want an AI that operates a computer rather than a robot, NorthCUA does the same idea on a desktop, free and open source, and it runs on hardware you already own.

What is still unclear

Three things in this story need a caution flag.

First, the 2.2 million downloads. That figure comes from Ant Group's own November 3 announcement at the Wuzhen Summit and covers the whole LingBot series, not version 2.0 alone. The public mirrors tell a smaller story: Hugging Face shows about 1,478 downloads for the v2-6b checkpoint, so most of the claimed downloads flow through ModelScope and other channels that are hard to check from outside China. The number is plausible for a series that has been out since January, but it is company-reported and not independently verifiable.

Second, the star count. Some coverage of the Wuzhen announcement repeated a figure of 33,000 GitHub stars for the project. The live repositories show 1,098 stars for lingbot-vla-v2 and 1,881 for the original lingbot-vla, checked on November 5, 2026, or about 3,000 combined. The larger figure does not match GitHub, so do not repeat it.

Third, real-world proof. Robbyant says the model is in commercial pilot testing with hardware partners Leju and Ti5 Robot and customers including GuoDa Drugstore and Longsheng Technology, in retail sorting, logistics, and industrial automation. Those pilots are company-reported, with no public results, and no independent evaluation of LingBot-VLA 2.0 itself has been published. Free open weights with a catch is a pattern in this space. Deep Principle's Hydrogen-1.0 is free too, but running it carries a $53,280 compute catch, and here the catch is the robot, the GPU, and the engineering time.

Who should actually use it

LingBot-VLA 2.0 makes sense for three groups. Robotics researchers get a strong, permissively licensed baseline they can fine-tune and publish against. Hardware makers get a ready-made brain that already speaks 20 robot configurations, which is cheaper than building a policy from scratch. Teams running pilot deployments in warehouses or stores get a starting point with documented commercial pilots behind it.

It does not make sense as a first robot project or a purchase decision: there is nothing to buy, and the model will not make a cheap robot reliable. If your team already has robots and GPUs, this is a strong free option. If you do not, the download count is not a reason to start.

Is LingBot-VLA 2.0 really free?

Yes, for the software. The code and the 6B model weights are under the Apache 2.0 license, which allows commercial use with no fee. The costs are the robot, a high-end GPU for real-time inference, and the engineering time to set it up and fine-tune it.

What robots can LingBot-VLA 2.0 control?

Twenty configurations spanning single-arm, dual-arm, bipedal, and wheeled robots from 17 manufacturers, including Unitree, AgileX, Galaxea, Franka, and Astribot. The model maps each machine into one shared 55-dimensional action space, so the same trained policy can drive any of them.

How fast does LingBot-VLA 2.0 run?

About 130 milliseconds per inference call on an NVIDIA GeForce RTX 4090D with 10 denoising steps, according to Robbyant's deployment guide. That is fast enough for real-time closed-loop control on a capable GPU.

How does LingBot-VLA 2.0 compare with π0.5?

On Robbyant's own GM-100 benchmark, LingBot-VLA 2.0 leads π0.5 on both tested platforms, 66.2 versus 59.1 progress score on the AgileX Cobot Magic. Both are open-weight and Apache 2.0 licensed, but π0.5 has the larger community, and no independent study has tested LingBot-VLA 2.0 head to head.

Can I run LingBot-VLA 2.0 at home?

Only if you have a compatible robot and a high-end GPU. The model itself is a free download, but it needs physical hardware to do anything, plus a Linux setup with Python 3.12 and PyTorch 2.8. For most people it is a project to watch, not one to install.

Ant Group's robot brain is free, open, and ahead of its open rivals on the vendor's own tests, yet still a research-grade policy that fails most tasks it has never seen. That combination is exactly what "open source" is for: anyone can check the claim, run the model, and find out. The 2.2 million downloads say a lot of people are doing exactly that. If you have the hardware, the download is worth your time. If you do not, the honest move is to wait for the independent results.

Sources

• Robbyant — LingBot-VLA 2.0 repository (official): github.com/Robbyant/lingbot-vla-v2
• Wei Wu et al. — "From Foundation to Application: Improving VLA Models in Practice," arXiv:2607.06403, submitted July 7, 2026: arxiv.org/abs/2607.06403
• Ant Group via GlobeNewswire — open-source presentation at the Wuzhen Summit, November 3, 2026: menafn.com (GlobeNewswire release)
• Hugging Face — robbyant/lingbot-vla-v2-6b model page: huggingface.co/robbyant/lingbot-vla-v2-6b
• MarkTechPost — "Robbyant Releases LingBot-VLA 2.0," July 8, 2026: marktechpost.com
• Tech Times — "Ant Group's Open-Source Robot Brain Beats pi0.5 Across 20 Hardware Types," July 11, 2026: techtimes.com
• AI Reporter America — "Ant Group's Robbyant Released LingBot-VLA 2.0 as an Open-Source Embodied AI Model": aireporteramerica.com
• Techietory — "Robot Foundation Models Compared: GR00T vs. π vs. Gemini Robotics vs. Open VLAs (2026 Guide)": techietory.com
• RoboColosseum — "Evaluating Physical Intelligence at Scale In-The-Wild," arXiv:2611.02497, November 2, 2026: robotpapers.info

Note: LingBot-VLA 2.0 is a research model for physical robots. The benchmarks in this article are company-reported unless stated otherwise, and the model has no published independent evaluation yet.