A picture generates a never-ending 3D world: LingBot-World 2.0 is open source today, with 60fps unbounded interaction

On the World Model track, someone dropped another depth charge. On July 9, the Ant Lingbo team (Robbyant) officially released LingBot-World 2.0 (codenamed LingBot-World-Infinity) as open source. The three-piece set of papers, code, and model weights were all made public. This time, they took down the two most difficult bones of the world model: "infinite duration" and "randomly changing world".

Let’s talk about the most explosive numbers first: 720p, 60fps, unlimited duration. It’s not “near infinite”, it’s true unbounded interaction under the causal pre-training paradigm – as long as your GPU is powered off, the model can continue to be generated, and the output quality will not deteriorate over time. Compared with version 1.0, which can only maintain stable generation for nearly 10 minutes, this generation is completely full.

Four core upgrades: from “usable” to “fun”

LingBot-World 2.0 brings qualitative changes in four dimensions, each of which directly hits the pain points of the world model.

① Unbounded interaction duration

This is the most hardcore breakthrough. Although version 1.0 supports stable generation in the past 10 minutes, there is an upper limit after all. 2.0 adopts a carefully designed causal pretraining paradigm, and the model can achieve unlimited interaction duration while maintaining consistent output quality. To put it simply: your character can keep running in the world, the picture will not become blurry, and the physical logic will not become more and more confusing.

The technical key lies in the sliding window design of KV Cache. The model does not stuff all historical frames into the video memory, but retains the sink frame (anchor frame) + the KV cache of the most recent N frames, moving forward like a flowing river. It not only ensures long-term consistency, but also controls video memory consumption.

② 60fps real-time response

Version 1.0 is 16FPS and the delay is within 1 second. 2.0 directly distills the real-time version to 720p 60fps. This is not a pre-rendered video playback, this is truly real-time generation - you press a button and the picture responds immediately. The team used model distillation technology to extract a lightweight real-time variant from the 14B basic model, which greatly reduced the inference delay while maintaining picture quality.

③ Richer interactive elements

Version 1.0 mainly supports camera control (translation, rotation) and basic actions (forward, jump). 2.0 greatly expands the action set: attack, archery, spell casting, and shooting are all supported. What's more important is the introduction of a text-driven event system - you can use text descriptions to change the state of the world, such as "summon a chair" or "conjure a flower", and the model will generate corresponding content in real time and integrate it into the scene.

④ Agentic Harness: AI becomes its own director

This is the most cutting-edge attempt. The team integrated an agent framework into the world model, including two roles:

• Pilot Agent: Responsible for planning and executing character behaviors, deciding where to go and what actions to take.
• Director Agent: Responsible for synthesizing new environmental elements and dynamically generating new content as the scene progresses.

It is equivalent to saying that the model itself can be used as a game engine + game planner. You only need to give it a picture and a text description, and it can run out of an evolving world on its own.

Model specifications: 14B flagship + 1.3B lightweight

Two sizes of models are released this time:

• 14B Ultimate Edition: Provides three variants of causal-fast (causal fast reasoning), causal-pretrain (causal pre-training), and bidirectional (two-way), with full performance
• 1.3B lightweight version: Provides two variants: causal-fast and causal-pretrain, supporting single-GPU deployment.

The 14B version requires 8 GPUs to run multi-card parallel inference, while the 1.3B version can run on a single card. Both versions have been uploaded HuggingFace and ModelScope. 1.0 is the Apache 2.0 loose license, and 2.0 is changed to CC BY-NC-SA 4.0 (non-commercial).

How to run

The code is built based on Wan2.2, and the installation process is very standard:

git clone https://github.com/robbyant/lingbot-world-v2.git
cd lingbot-world-v2
pip install -r requirements.txt
pip install flash-attn –no-build-isolation

Download model weights (taking 14B causal-fast as an example):

huggingface-cli download robbyant/lingbot-world-v2-14b-causal-fast –local-dir ./lingbot-world-v2-14b-causal-fast

Run a 480P causal inference:

torchrun –nproc_per_node=8 generate.py \
–task i2v-A14B –size 480*832 \
–ckpt_dir lingbot-world-v2-14b-causal-fast \
–image examples/03/image.jpg \
–action_path examples/03 \
–dit_fsdp –t5_fsdp –ulysses_size 8 \
–frame_num 361 –local_attn_size 18 –sink_size 6 \
–prompt “A serene lakeside scene with a lone tree standing in calm water…”

Key parameter description:
– frame_num: The number of generated frames, 361 frames takes about 22 seconds @16fps, but it can be set larger under causal reasoning.
– local_attn_size: Local attention window size, controls how many historical frames are viewed at each step
– sink_size: The number of anchor frames, retaining the first few frames as global context

Online trial channel

Don’t have an 8-card A100? It doesn’t matter, the team provides two online experience platforms:

• Overseas (Reactor)：reactor.inc/lingbot-world-v2, play directly on the web page
• Domestic (Emmanuel)：lingguang.com/support, available on mobile phone

In addition, there will be an official complete Demo booth at WAIC 2026 (World Artificial Intelligence Conference) at the end of July. Those who want to experience the full-blooded version can go there then.

Deployment plan: SGLang has native support

If you plan to deploy it yourself, SGLang already natively supports the real-time inference path of LingBot-World 2.0. It uses the WebSocket + MessagePack protocol. The server maintains an active session, generates video frames chunk by chunk, and supports dynamically receiving control events (camera actions, text prompt changes, etc.) at runtime.

The core API is /v1/realtime_video/generate WebSocket endpoint, after the client connects, it first sends an init message (including the first frame image, prompt, size, FPS and other parameters), and then can continue to send event messages to update the scene. Supported event types include prompt update, camera action, and composite input.

How does it compare to Google Genie 3?

When version 1.0 was released, the team clearly targeted Google Genie 3. From 1.0 to 2.0, the gap has not only been equalized, but also surpassed in some dimensions:

• duration: Genie 3 limited time vs LingBot 2.0 unbounded
• Resolution/FPS: Genie 3 undisclosed detailed specifications vs LingBot 2.0 clear 720p@60fps
• Open source level: Genie 3 closed source vs LingBot 2.0 fully open source (paper + code + weight)
• interaction richness: LingBot 2.0’s Agentic Harness is the first in the world model field

limitations and deficiencies

Objectively speaking, there are a few limitations:

• The 14B version requires an 8-card GPU, and the threshold is not low (the 1.3B version alleviates this problem)
• 2.0 license changed to CC BY-NC-SA 4.0, commercial use requires separate authorization
• The team made it clear that it will not release the deployment code and needs to refer to SGLang or flashdreams to do it.
• The real-time 60fps version is a distilled lightweight variant, and the image quality is theoretically not as good as the full health 14B

what does it mean

World models are moving from "laboratory demo" to "usable product". The three key breakthroughs of LingBot-World 2.0 – unlimited duration, real-time response, and agent-driven – exactly correspond to the core needs of the three major application scenarios of games, virtual worlds, and robot training.

Imagine: with a picture + a piece of text, AI can generate a 3D world that you can freely explore and never stop. There are characters who can cast spells, new objects that appear randomly, and a continuously evolving environment. This is no longer science fiction, this is code that can be run today.

The Ant Lingbo team has built a complete embodied intelligence technology stack from LingBot-Depth (spatial perception) to LingBot-VLA (embodied decision-making) to LingBot-World (world simulation), plus the newly open source LingBot-Map (real-time mapping). The world model is the most critical piece of the puzzle - it allows AI not only to "understand" the world, but also to "create" the world.

A picture, a sentence, an infinite world. This is the world model of 2026.

Project address: github.com/Robbyant/lingbot-world-v2
Paper: arxiv.org/abs/2607.07534
Online trial: reactor.inc/lingbot-world-v2