蚂蚁灵光App上线世界模型功能：LingBot-World-Fast开源，手机端一图生成3D世界

The Ant Lingguang App has launched the "Experience World Model" feature today, integrating the open-source model LingBot-World-Fast. By uploading a single image, users can generate and explore a 3D world directly on their mobile devices — marking the first time a world model has been implemented on mobile.

One Image, One World — Running on Your Phone

On April 27, Ant Group’s multimodal AI assistant Ling Guang officially launched the “Experience World Model” feature. Simply put: you give it an image, and within seconds, it generates a 3D world you can explore from a first-person perspective.

Sounds like a flashy demo? Not entirely. The key this time isn’t “another world model,” but rather “it actually runs on your phone.”

Over the past year, world models have been one of the hottest concepts in the AI community. From Sora’s video generation boom to academic projects like Genie and UniSim, and to internal experiments by various embodied AI teams — everyone has been trying to make AI not only “understand” the world but also “simulate” it. Yet most world model experiences are limited to: opening a web demo, waiting tens of seconds or minutes, then watching a pre-rendered video.

Ling Guang has done something different — it has compressed the inference chain of the world model down to a level usable on mobile devices, supporting real-time interaction. Users don’t watch a video; they play, just like a mobile game, using a virtual joystick to move forward and backward, turn freely, and “walk around” inside an AI-generated 3D environment.

This difference in experience is much more tangible than a few extra billions of parameters.

How to Play: Three Steps to Enter “Your World”

The process is intuitive:

Open the Ling Guang app and upload a picture in the dialogue box.
The system will smartly recommend an option — choose “Generate world from image,” or directly enter a command such as “Explore this world in first-person view.”
Wait a few seconds for the 3D world to be generated, then click to enter.

After entering, a virtual joystick appears on the screen. The controls are just like in a mobile game — the left side moves, and the right side (or screen swipe) changes the perspective. Each session allows up to 60 seconds of continuous exploration.

Sixty seconds may not sound long, but considering this is a real-time generated and rendered 3D scene — not a pre-recorded video — it’s quite impressive. More importantly, Ant Group officially noted that the model supports “minute-level long-term consistency” — meaning if you walk around and return to the starting point, the buildings, vegetation, and lighting around you won’t suddenly change. That’s a hard technical benchmark for world models — one where many academic demos have stumbled badly.

The Model Behind It: LingBot-World-Fast

The new feature in Ling Guang is powered by Ant’s LingBot team’s LingBot-World-Fast world model.

First, some background on the LingBot team. LingBot is Ant Group’s core research team focusing on embodied intelligence and multimodal AI. This year, they have open-sourced several models:

LingBot-Depth – a high-precision spatial perception model addressing “how AI understands 3D space.”
LingBot-VLA – an embodied large model enabling AI to make motion decisions based on environmental understanding.
LingBot-World – the world model, this feature’s centerpiece.

Together, these three models form a complete chain: “perceive space → simulate world → make decisions.” LingBot-World handles the middle part — given an initial observation (e.g., an image), it predicts “what the world will look like if I walk three steps forward and turn right 90 degrees.”

LingBot-World-Fast is a lightweight, accelerated version of LingBot-World, optimized for real-time interactive scenarios. Based on naming conventions, it likely involves architectural pruning and inference acceleration for mobile-level latency performance.

Notably, LingBot-World-Fast is open-source, meaning developers can run, modify, or integrate it into their own applications. For teams working on gaming, XR, or embodied AI, this is a practical starting point.

What Problem Does a World Model Solve?

Let’s take a step back and discuss what “world models” actually mean.

In the simplest terms: a world model is an AI that can “imagine” how the world operates. Given a current state, it can predict the next one. Conceptually, it’s the same as a large language model predicting the next token — except the predictions are visual signals in 3D space rather than text.

Why is this important? At least three reasons:

1. The “Brain Simulator” for Embodied Intelligence

Robots acting in the real world can’t rely purely on trial and error. They need an internal model to “imagine” the consequences of actions and choose the optimal strategy. The world model serves as this internal simulator. The LingBot team’s pipeline linking Depth → World → VLA clearly reflects this thought process: first understand space, then simulate the world, finally make decisions.

2. The Next Generation of Content Creation

From text generation to image generation to video generation, the next natural step is interactive 3D scene generation. A capable world model could replace much of the manual modeling in traditional game engines. Feed in a concept image, and get a walkable environment — highly attractive for gaming, film, and architectural visualization industries alike.

3. A Necessary Step Toward AGI

Yann LeCun repeatedly emphasizes that world models are a key component in reaching AGI. His argument: a truly intelligent system must understand how the physical world operates, not just perform language-level reasoning. Whether or not you agree with his technical approach, world models undeniably represent a vital path for AI to comprehend the physical world.

What’s Hard About Bringing It to Mobile?

Building a world model is one thing — making it run on a phone is another.

The inference process for a world model is essentially conditional video generation — given the current frame and user action, generate the next frame(s). This is computation-heavy, traditionally requiring high-end GPU servers.

To make real-time mobile interaction possible, at least three challenges must be solved:

Latency: The response time between joystick movement and screen update must be under a few hundred milliseconds, or the experience feels dizzying. Models must infer extremely fast or use predictive generation (precompute several possible viewpoints).
Consistency: Sequentially generated frames must maintain spatial continuity. If you walk forward 10 steps then turn back, the scenery behind shouldn’t have morphed. This is where world models often fail — especially in long sequences.
Quality vs. compute balance: Mobile devices have strict performance and thermal limits. Maintaining acceptable graphics under these constraints requires highly efficient modeling.

Ant’s official notes say LingBot-World-Fast supports “minute-level long-term consistency” and “real-time interaction.” If true, that implies significant progress in model compression and inference optimization. Still, “real-time” (frame rate? resolution?) and “consistency” (actual stability) are open questions for developers to validate through hands-on testing.

Strategically, the 60-second limit is a pragmatic choice — enough for users to feel the novelty of world models without overloading the device or breaking consistency over longer periods.

How Ling Guang Stands Out

Currently, public world model products are rare. Google DeepMind’s Genie exists mainly in academic papers and limited demos; Decart’s Oasis offers a Minecraft-like generative game running on PCs; World Labs raised major funding but hasn’t released a product.

Ling Guang’s differentiation lies in two aspects:

1. Native mobile experience. Not “a web demo opened in a phone browser,” but a feature deeply integrated in the app, optimized for touch controls. This isn’t just engineering — it defines the user acquisition path. Playing casually on mobile versus setting up a PC demo is a night-and-day difference.

2. Model open-source. By open-sourcing LingBot-World-Fast, this is not a closed feature but an open technical capability. Developers can extend and integrate it freely — crucial for nurturing the ecosystem of an emerging field.

Of course, limitations remain: 60-second sessions, single-image inputs, and unclear visual fidelity all indicate this is more of a technical preview than a finished product. But as the world model’s first public mobile deployment, its symbolic value outweighs its product maturity.

Open-Source Ecosystem: LingBot’s Growing Puzzle

Looking broadly, Ant’s LingBot team’s open-source strategy shows a clear rhythm.

From LingBot-Depth to LingBot-VLA to LingBot-World, they didn’t publish one massive all-in-one model, but rather incrementally open-sourced along the chain of “perception → simulation → decision.” Each model solves a focused subproblem, yet together they form a complete embodied intelligence pipeline.

This modular open-source approach offers several benefits:

Each model can be used independently, lowering the entry barrier for developers.
Modules can be individually improved or swapped by the community, increasing flexibility.
Gradual release pacing gives the team room to iterate and refine.

For developers interested in embodied AI or world models, the LingBot series is already a project worth long-term attention.

World Models in 2025–2026: The Transition from Paper to Product

Looking back over the past year, progress in world models has clearly accelerated. Early 2024 saw Sora’s launch make video generation a household topic, but true world models — those understanding physics and supporting interactive exploration — still largely existed in academic papers.

By late 2025, the landscape had shifted. Multiple teams released interactive demos, and academic metrics like long-term consistency and physical accuracy improved significantly. Now, in April 2026, Ant’s Ling Guang has brought a world model to mobile — and open-sourced the underlying engine.

That progress has moved faster than many expected.

Of course, there’s still a long way to go before world models become truly “usable.” In complex scenes, they still struggle with physical consistency, long-run stability, and fine-grained rendering quality. But the direction is clear, and the path to deployment increasingly visible.

Ling Guang’s update is less a product launch than a public milestone — proof that world models have moved from “lab papers” to functional features on smartphones. The distance between those two points is shorter than most imagined.