LingBot-Video: open-source MoE video generation model

LingBot-Video is an open-source large-scale MoE video generation model dedicated to embodied intelligence.

Verified

Core facts

Read by a script from the official API, not rewritten by a model.

License
Apache-2.0
Languages
Python
Maintainer
Robbyant
Verified

What it is

LingBot-Video is an open-source large-scale MoE (Mixture-of-Experts) video generation model dedicated to embodied intelligence. LingBot-Video is designed to bridge the gap between video synthesis and physical world understanding. This project is licensed under the Apache 2.0 License.

Main features

  1. The project provides DMD-distilled checkpoints (t2v and ti2v) for 8-step video generation, on both the diffusers and SGLang engines.

  2. The project says the model is trained on massive web videos integrated with at least 70,000 hours of embodied data.

  3. The project says the model is rewarded for high aesthetics, physical rationality and task completion.

  4. LingBot-Video-Dense has the component Dense (1.3B) and the tasks T2I, T2V and TI2V.

  5. LingBot-Video-MoE has the components MoE (30B-A3B) plus Refiner and the tasks T2I, T2V, TI2V and Refinement.

  6. LingBot-Video-MoE-DMD has the component MoE (30B-A3B) as an 8-step distilled student and the tasks T2V and TI2V.

Getting started

  1. The root requirements file includes the recommended PyTorch build for LingBot-Video inference.

  2. Installing the SGLang requirements file enables the SGLang runtime.

  3. Among the runtime versions, Python is listed at 3.10 or newer.

  4. The runtime versions also list transformers 5.8.1, diffusers 0.39.0 and peft 0.19.1.

  5. The project provides ready-to-run scripts for single-GPU and multi-GPU inference.

  6. Single-GPU scripts use direct diffusers and sequential CFG by default to reduce peak memory.

  7. The diffusers backend is the direct diffusers reference path, and the sglang backend is the SGLang Diffusion path.

From the README · bash
git clone https://github.com/Robbyant/lingbot-video
cd lingbot-video

python -m venv .venv
source .venv/bin/activate
python -m pip install -U pip

# Base requirements cover direct DiT inference and rewriter --backend transformers.
pip install -r requirements.txt
pip install -e .
From the README · bash
python -m pip install --no-deps -r requirements-sglang.txt
From the README · bash
source .venv/bin/activate
export PYTHON_BIN=python
export DENSE_MODEL_DIR="<path_to_lingbot-video-dense>"
export MOE_MODEL_DIR="<path_to_lingbot-video-moe>"

Who it is for

LingBot-Video DiT inference is designed to consume structured JSON captions. LingBot-Video-Rewriter-Base uses Qwen3.6-27B official for Prompt rewriter (Expand). LingBot-Video-Rewriter-Adapter uses Qwen3.6-27B LoRA for Prompt rewriter (Json).

Multi-GPU refiner scripts use CP8 + FSDP + sequential CFG by default. MoE scripts default to grouped expert execution. The project's internal evaluation assesses T2V and TI2V capabilities for open-source models across two dimensions, Quality Scores and Domain Scores.

On Sept 18, 2026, the project released the 8-step DMD student model for LingBot-Video. On Aug 5, 2026, the project released the RBench evaluation for LingBot-Video. On July 9, 2026, the project released the technical report, code, models and rewriters for LingBot-Video.

Sources

This page is written from the official sources below. Open them to check the original text.

Verified ·

This page is written by AI from the official sources above and checked automatically. This site has not tested it. Found a mistake? Tell us.