Ant Group's LingBot-World generates up to 10 minutes in one pass
The open-source world model runs interactively at 16 frames per second with sub-second latency, and its maker published the code and weights under Apache 2.0.

Key takeaways
- Robbyant, the embodied-AI company inside Ant Group, announced the open-source release of the world model LingBot-World on January 29, 2026, publishing the code and model weights under the Apache 2.0 license.
- Robbyant says LingBot-World generates up to nearly 10 minutes of continuous video in one pass and runs interactively at about 16 frames per second with latency under one second.
- The technical report puts LingBot-World at 28B parameters split into two 14B experts, with only one expert active at each denoising step.
- The same report lists the model's limits: memory is emergent from the context window, so long runs drift, and the model needs enterprise-grade GPUs.
This story covers a release from 29 January 2026. It was published on 18 November 2026; nothing in it is new as of that date.
Ant Group's embodied-AI unit has open-sourced a world model that generates up to 10 minutes of coherent video in a single pass, then lets you steer what happens next. The unit, Robbyant, announced the code and weights for LingBot-World on January 29, 2026, and its technical report describes the longest single generation it claims for an open model of this class.
A world model predicts what a scene does next rather than rendering one fixed clip: objects stay in place when the camera looks away, the layout holds, and the view changes when you press a key. LingBot-World is built to show all three at once.
What Is LingBot-World?
LingBot-World is an open-source world simulator that turns one image and a text prompt into an interactive video stream. Robbyant, the embodied-intelligence company inside Ant Group, announced the open-source release of the code and model weights on January 29, 2026. The GitHub repository and the Hugging Face model cards both carry that release date, and the technical report, arXiv:2601.20540, posted a day earlier, on January 28, 2026.
The date has been muddled in secondary coverage. A community summary that recirculated in November dated the release to November 17; Robbyant's own press release and repository place it on January 29, 2026, while the technical report went up on January 28.
Robbyant ships three variants. LingBot-World-Base (Cam) takes camera-pose control, LingBot-World-Base (Act) takes action commands, and LingBot-World-Fast is tuned for low-latency interaction, the project's documentation says. All three generate at 480P and 720P. The codebase is built on Wan2.2, and the project ships under the Apache 2.0 license.
How Does LingBot-World Generate 10 Minutes in One Pass?
It starts from a 14B-parameter image-to-video diffusion model and is trained in stages to hold a scene together far longer than a normal clip. Robbyant's report says the base is Wan2.2's 14B image-to-video model. Training runs in three phases: pre-training for general video quality, middle-training that adds world knowledge and action control, and post-training that converts the model for real-time use.
The size figure comes from the same report. The middle-trained model uses a mixture-of-experts design, with a high-noise expert for coarse layout and a low-noise expert for fine detail. Each expert holds about 14B parameters, for 28B in total, and only one runs at a time, so inference cost stays close to that of a dense 14B model, the report states.
For the long horizon, the report says the team adapted the model to causal attention and distilled it into a few-step autoregressive system, which is what brings latency under a second. Long runs spread across multiple GPUs with FSDP and DeepSpeed Ulysses, per the model card. Robbyant's headline figure is "up to nearly 10 minutes of continuous, stable, and lossless video generation" in one pass, from the release announcement. The report's own wording is "up to ten minutes."
What Does LingBot-World Need to Run?
Enterprise hardware, not a desktop. The report lists high inference cost as a limit and says the model needs enterprise-grade GPUs. The repository's example commands run across eight GPUs, and the model card notes that a larger frame count, roughly 961 frames for a minute at 16 frames per second, requires enough CUDA memory. The release offers no consumer-GPU path.
Where LingBot-World Still Fails
Robbyant's own limitations section is direct about the gaps. The report says the model's memory "is an emergent ability derived from the context window rather than an explicit storage module," so it lacks stability and drifts over long simulations. It also flags the inference cost, an action space limited mostly to navigation and basic movement, trouble with fine-grained object interaction such as picking up a specific cup, and support for only a single agent. The project page adds that real-time performance comes at the cost of some visual fidelity, and that the roadmap points to an explicit memory module and a wider action space.
How Does LingBot-World Compare?
On the benchmark its maker chose, LingBot-World leads the two open models it was measured against. The report compares it with Yume-1.5 and HY-World 1.5 on VBench. Across six metrics, LingBot-World scores highest on imaging quality (0.6683, against 0.5838 and 0.6512), aesthetic quality (0.5660), dynamic degree (0.8857, against 0.7612 and 0.7217) and overall consistency (0.2178), and trails HY-World 1.5 on motion smoothness and temporal flickering, according to the report's table. The figures are the maker's own, measured on 100 generated clips longer than 30 seconds each.
The closer comparison in the field is Google's Genie 3, which the report places in the same long generation horizon, though its table gives LingBot-World the higher dynamic-degree class, and which is not open-source. The report says LingBot-World is the only high-capability model in its comparison that is fully open-sourced. Community coverage went further: a summary at himiao.ai, drawing on InfoQ's write-up, called the near-10-minute run a likely record for a single generation among current video and world models. The same summary describes the demo as a 9-minute-20-second first-person walk, from a ruined ancient Greek temple through city paths to a neoclassical building complex, with no editing or splicing, and notes flaws in the final minutes when the spatial relationship between buildings is briefly forgotten.
What Comes Next for LingBot-World?
The line has already moved on. Robbyant's original LingBot-World repository is no longer actively maintained; a banner on the GitHub page points to LingBot-World 2.0, also called LingBot-World-Infinity, which the team released on July 9, 2026. The 2.0 report describes an unbounded interaction horizon, 720p streams at 60 frames per second, a wider set of actions, and an agentic harness, and the release carries a non-commercial CC BY-NC-SA 4.0 license, stricter than the Apache 2.0 terms on version 1.0.
Robbyant shipped two other embodied models in the same week as LingBot-World 1.0: the spatial-perception model LingBot-Depth and the embodied foundation model LingBot-VLA, the company said in the release that announced LingBot-World.
The code and weights for LingBot-World 1.0 are on GitHub, Hugging Face and ModelScope under Apache 2.0, but the repository directs new work to the LingBot-World 2.0 release.