Chengtao Lv

🇸🇬 PhD @NTUsg | Efficient AI

[Chengtao Lv @mackWilliams319]

awesome!

[Mako AI @LightX2V]

15s of MiniMax-H3 video + audio, generated in ~72s on one RTX 5090.
Our baseline took ~26 minutes. We wrote up how LightX2V got it down to 72s 👇
768×1344, steady-state benchmark. MP4 export and network transfer excluded. https://t.co/0sMSdJfcIQ

[Chengtao Lv @mackWilliams319]

@QuanyuLoong Great work!

[Quanyu Long @QuanyuLoong]

🤔 What if various environments became interactive LLM agents?

🚀 Introducing Trace2Env: a training-free framework for agentic language world modeling, supporting stateful and long-horizon environment simulation.

🕹️ Below is a “fake” Terminal, observations of it are purely generated.

#WorldModels #AIAgents

[Chengtao Lv @mackWilliams319]

🚀 Excited to release RealtimeWAM! The first one-step asynchronous World Action Model with superior task performance.

⚡ ~12 ms action generation with a 6B model on a single H100 GPU — ~25× faster! ⚡

✨ Supports any MoT-based WAM (e.g., Fast-WAM and Faster-WAM), with less than 1% success-rate loss on diverse benchmarks (e.g., RoboTwin 2.0 and LIBERO-Plus).

🧩 Algorithm–system co-design: few-step distillation, blockwise asynchronous inference, and efficient kernels—all implemented in LightX2V.

🙌 Welcome to check it out, try it, and share feedback!

📄 Paper: https://t.co/aE80M7nM9H
🤗 Model weights: https://t.co/qzpOuBNbk1
⭐ Code: https://t.co/h6KnAkbkiZ