LingBot-World-V2 generates endless interactive video worlds, under a non-commercial licence

What a world model means in this project

LingBot-World 2.0, also called LingBot-World-Infinity, generates video that keeps responding to input rather than playing back a fixed clip. You give it a starting image and a stream of actions, and it continues the scene. The README frames four upgrades over the first version.

Unbounded interaction horizon, meaning the model keeps producing consistent output over an arbitrarily long exchange, attributed to a causal pretraining paradigm. Rapid response time, achieved by distilling a real-time variant from the base model, described as sufficient to drive 720p video at 60 frames per second. A wider set of interactive elements, with the README naming attacking, archery, spell-casting and shooting, plus a richer variety of text-driven events. And an agentic harness, which the authors present as the first integration of that kind inside world modelling.

The audience is researchers and engineers working on interactive video generation or embodied simulation. There is a hosted way to try it without hardware, covered below, but the interesting material is the code and the weights.

Four checkpoints, and what the names tell you

Four model variants are published. The 14B causal-fast checkpoint is the distilled real-time one. The 14B causal-pretrain checkpoint is the pretrained causal model. The 14B bid checkpoint is the bidirectional variant. And there is a 1.3B causal-fast checkpoint for a much smaller footprint.

The distinction that matters operationally is causal versus bidirectional. Causal inference processes frames chunk by chunk with a KV cache instead of all at once, which is what makes an unbounded horizon possible at all. The trade is that a bidirectional model sees the whole sequence and can produce a more coherent fixed-length clip, so the two are not substitutes.

One packaging detail will cost you an hour if you miss it. The 1.3B HuggingFace package contains the DiT weights only. T5, the VAE and the tokenizer are shared with the 14B release and have to be passed in separately, either through an assets directory argument or as the third argument to the helper script.

All four are on HuggingFace, and the 14B causal-fast checkpoint is also on ModelScope.

Two agents, one driving the character and one driving the scene

The agentic harness is the most original part of the release and the part with the least documentation, which is worth saying plainly.

As described, a pilot agent plans and executes character behaviours, while a director agent synthesises novel environmental elements as the scene progresses. That split maps onto the two things a long interactive generation has to keep doing: respond to what the user just did, and keep the world from running out of ideas.

What the README does not provide is the harness code, its prompts, or its evaluation. The repository tree contains the generation script, a helper script, the wan package, examples and a paper, and the TODO list marks all four checkpoint releases as done without mentioning the harness. Treat the harness as a described result rather than something you can run from this repository.

That gap matters if your interest is the interaction loop rather than the video model, because it is the part you would have to rebuild.

Installing it and getting the weights

The codebase is built on Wan2.2 and the README directs you to that project's documentation for installation. Python 3.10 or newer is required, and the dependency set is a conventional diffusion stack with a few hard pins, including torch at 2.4.0 or newer, transformers capped at 4.51.3, and numpy held below 2.

Clone and install:

git clone https://github.com/robbyant/lingbot-world-v2.git
cd lingbot-world-v2

# Ensure torch >= 2.4.0
pip install -r requirements.txt

Flash attention is installed separately, and needs the no-build-isolation flag:

pip install flash-attn --no-build-isolation

Weights come down through either the HuggingFace CLI or the ModelScope CLI. The HuggingFace route is the one that covers all four variants:

pip install "huggingface_hub[cli]"
huggingface-cli download robbyant/lingbot-world-v2-14b-causal-fast --local-dir ./lingbot-world-v2-14b-causal-fast
huggingface-cli download robbyant/lingbot-world-v2-1.3b-causal-fast --local-dir ./lingbot-world-v2-1.3b-causal-fast/transformers

If HuggingFace is slow or blocked where you are, the ModelScope route is documented for the 14B fast checkpoint:

pip install modelscope
modelscope download robbyant/lingbot-world-v2-14b-causal-fast --local_dir ./lingbot-world-v2-14b-causal-fast

The repository ships a Makefile with a format target running isort and yapf, and a pyproject marking version 1.0.0, so this is a packaged research release rather than a notebook dump.

Running inference, and the GPU arithmetic

Generation goes through a script that performs causal inference with a KV cache, processing frames chunk by chunk. The README gives three example invocations and one helper script.

The pattern to internalise is that GPU count and sequence-parallel size are tied to the model. The 14B causal-fast example runs on eight GPUs at 480 by 832, with a sequence-parallel size of eight that must divide the model's forty attention heads. The 1.3B causal-fast example runs on four GPUs at the same resolution, with a sequence-parallel size of four that must divide twelve heads. The causal-pretrain example also uses eight GPUs but a much shorter frame count.

There is a shell helper that infers both the task and the GPU count from the name of the checkpoint directory, so a directory containing the 1.3B marker is treated as the small model on four GPUs and anything else as the 14B on eight. It takes the weights directory, a frame count, and optionally the shared assets directory.

So the honest hardware requirement is a single node with four or eight GPUs, depending on which checkpoint you pick. There is no single-GPU path documented, and no quantisation or offloading guidance in the README.

The deployment gap and the hosted demos

The most consequential sentence in the README is under Deployment: the authors state they do not plan to release their deployment code. If you want to serve this yourself at interactive latency, they point you at the LingBot-World deployment in SGLang or at NVIDIA's flashdreams.

That is a real fork in the road. Research code that runs one generation at a time and production serving code that sustains sixty frames per second are different engineering problems, and this repository deliberately stops at the first.

For hands-on evaluation without hardware, two hosted options are listed: Reactor for the international web demo and LingGuang for a domestic mobile experience. The README attaches a note that in the authors' own setup the model runs at full capability, and directs people who want the official demo to WAIC 2026, which is a fair hint that the hosted versions are not running the same configuration as the reference hardware.

Licence, and what share-alike costs you

The project is licensed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0. The README spells out the consequences: non-commercial use only, you may share and adapt with proper attribution, and derivative works must be distributed under the same licence.

Share-alike is the clause that surprises people. If you fine-tune these weights and release the result, the result has to carry CC BY-NC-SA 4.0 as well. For academic work that is usually fine. For anything that might later be commercialised, it is a one-way door, and worth settling before you start rather than after.

Note also that GitHub reports no recognised licence identifier for the repository, so automated tooling that filters by SPDX identifier may not classify it correctly.

Maintenance status is modest but not concerning: the last push was on 2026-09-10, matching the news entry for the remaining model releases on the same day, and the repository is not archived. There are no tagged releases, so you are tracking the default branch.

Against Wan2.2 and other routes to interactive video

The direct ancestor is Wan2.2, whose code and models this project builds on, and the README credits the Wan team explicitly. If what you need is high quality image-to-video generation without interaction, Wan2.2 is the simpler dependency and has a larger community around it. What LingBot-World-V2 adds is the causal formulation, the distilled fast variant, the action conditioning and the scene-level agent structure, which is exactly the set of things you need for a world rather than a clip.

On the interactive side, the 720p at 60 frames per second claim is the number to test rather than trust, because it is stated for the authors' own setup and the deployment code that would let you reproduce the serving stack is not published. That is the single biggest thing to verify if latency is your requirement.

The practical starting point is the 1.3B fast checkpoint on four GPUs if you have them, or the 14B fast checkpoint on eight if you want the configuration the README documents most thoroughly. Either way, budget for doing your own serving work.

Editorial conclusion

Download it if you are researching interactive video generation, have a multi-GPU node, and can work under a non-commercial licence: the checkpoints and inference code are real, the technical report is on arXiv, and the causal fast variant is the one to start with. Do not plan a product around it, since CC BY-NC-SA 4.0 excludes commercial use and requires derivative works to carry the same licence, and the authors state they do not intend to release their deployment code, so serving it at scale is work you do yourself or borrow from SGLang or flashdreams. Verify one thing before you commit the download bandwidth: whether the 1.3B route is viable for you, because its HuggingFace package contains only the DiT weights and needs the 14B release passed separately for T5, the VAE and the tokenizer.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. Robbyant/lingbot-world-v2 on GitHub