NVIDIA发了能在设备上实时跑的世界模型，4B参数，15Hz控制机器人，vLLM直接就能用。
NVIDIA 发布 Cosmos 3 Edge，一个4B参数的设备端世界模型，基于自研Nemotron骨干从零训练。模型采用双Transformer塔架构：自回归塔处理视觉/文本推理，扩散塔负责预测、生成和动作输出。在640×360分辨率下，单次推理可输出32个动作，于Jetson Thor上实现15Hz实时控制。vLLM和vLLM Omni已提供即日支持，可用于机器人、自动驾驶和视觉代理。
🚀 Day-0 support for NVIDIA Cosmos 3 Edge on vLLM a…
🚀 Day-0 support for NVIDIA Cosmos 3 Edge on vLLM and vLLM Omni!
An open on-device world model that lets robots, AVs, and vision agents reason over live video and generate robot actions locally.
🧩 4B params, built on two transformer towers over a shared representation: an autoregressive tower for vision/text reasoning + a diffusion tower for prediction, generation, and action 🧠 Built on a custom Nemotron backbone, trained from scratch — NVIDIA authored a dedicated vLLM integration to bring it up 🎮 Robot-control resolution (640×360), 32 actions/inference, real-time 15 Hz control on Jetson Thor
Serve it on vLLM and vLLM Omni today.