lingbot-video

Robbyant/lingbot-video is ranked 12,290 among these 8 peers; grammarly/gector is ranked 12,287.

As a top-tier video model, LingBot-Video is designed to bridge the gap between video synthesis and physical world understanding. 🔥 Key Highlights **🚀 Efficient MoE Architecture**: Scaled from scratch; balanced between capacity and cost with **~3x** faster inference. **📦 Data Engine**: Trained on massive web videos integrated with **70,000+ hours** of embodied data. **⚖️ Multi Reward System**: Rewarded for **high aesthetics**, **physical rationality**, and **task completion**. **⚡ Few-Step Student Sampling**: DMD-distilled checkpoints (t2v and ti2v) for 8-step video generation

, on both the diffusers and SGLang engines. 🎬 Video Demos 🔥 Latest News Sept 18, 2026: ⚡ We release the 8-step DMD student model for LingBot-Video. Aug 5, 2026: 🎉 We release the RBench evaluation for LingBot-Video. July 9, 2026: 🎉 We release the technical report, code, models, rewriters for LingBot-Video. 3B) | T2I, T2V, TI2V | 🤗 Huggingface 🤖 ModelScope | | 💪 LingBot-Video-MoE | MoE (30B-A3B) + Refiner | T2I, T2V, TI2V, Refinement | 🤗 Huggingface 🤖 ModelScope | | ⚡ LingBot-Video-MoE-DMD | MoE (30B-A3B), 8-step distilled student

| T2V, TI2V | 🤗 Huggingface 🤖 ModelScope | | 📝 LingBot-Video-Rewriter-Base | Qwen3. 6-27B official | Prompt rewriter (Expand) | 🤗 Huggingface 🤖 ModelScope | | 📝 LingBot-Video-Rewriter-Adapter | Qwen3. **💡 Rewriter deployment**: the bundled rewriter uses the single-process backend. For higher throughput, deploy the VLM yourself and call it through an OpenAI-compatible API. Preserve the two-stage semantics: step 1 must use the base VLM without the rewriter LoRA, while step 2 must use the same base VLM with the rewriter LoRA enabled. This can be

implemented with two endpoints, or with one server that can select the adapter per request. The default examples produce **121 frames at 832 × 480 and 24 fps**, without a refiner. must be a downloaded **local model root**, containing , , , , , and . Its must contain the DMD student weights. A transformer-only export is not sufficient. Start with the structured prompts and matching first frame already bundled in this repository: The scripts accept , , (TI2V), , , , , , and environment overrides. For example, to use your own

prepared TI2V input: Or call the unified runner directly: For TI2V, replace the mode and prompt with and add . For new plain-language prompts, follow Prompt Preparation and use the same first frame for rewriting and TI2V generation. 0, the examples do not need an Auto Negative step. The existing ordinary MoE scripts use a different scheduler and guidance recipe; use the DMD scripts above for student checkpoints. The single-GPU 5-second examples were checked on a GPU with

about 140 GiB of memory, using BF16 DiT/text encoder and FP32 VAE. Observed PyTorch peak allocated memory was about 82 GiB; this is not a minimum-memory specification. See the DMD guide for sampling and precision details. 🧭 Recommended Inference Workflow LingBot-Video DiT inference is designed to consume structured JSON captions, not casual natural-language prompts. Rewrite the user's plain prompt with Prompt Rewriter. For TI2V, pass the same first frame to the rewriter. Run Auto Negative by default

to prune the negative prompt for this specific caption. Run the unified inference runner with and select direct diffusers or