Deploy NVIDIA Cosmos Reason-1: Zero‑Code Physical AI on Alibaba Cloud PAI

Introduction

NVIDIA Cosmos is a World Foundation Model (WFM) development platform that accelerates physical AI for autonomous driving and robotics, providing advanced visual taggers, guardrails, and video data processing pipelines.

Cosmos Reason-1 Model

Cosmos Reason-1 is a fully customizable multimodal AI reasoning model designed to understand motion, object interaction, and spatio‑temporal relationships. It uses chain‑of‑thought (CoT) reasoning to interpret visual inputs, generate predictions from prompts, and reward optimal decisions. The model incorporates real‑world physics to produce context‑aware natural‑language responses and can serve as a discriminator or annotator for large‑scale synthetic data.

Cosmos Reason-1‑7B is fine‑tuned from Qwen2.5‑VL with physical‑knowledge and embodied‑reasoning data, employing supervised fine‑tuning (SFT) and reinforcement learning (RL).

Resources

NVIDIA Research: https://x.sm.cn/CO5McsA

NVIDIA Cosmos: https://x.sm.cn/DjS8lrQ

NVIDIA Cosmos Developer: https://x.sm.cn/1LXdBLx

PAI‑Model Gallery

The Alibaba Cloud PAI‑Model Gallery integrates Cosmos Reason-1, offering enterprise‑grade deployment with zero‑code setup, automatic cloud resource adaptation, ready‑to‑use APIs, and full‑process managed operations.

One‑Click Deployment Steps

Locate the Cosmos Reason-1‑7B model in the PAI‑Model Gallery or use the direct link https://x.sm.cn/IY41jvi.

Click “Deploy”, select compute resources, and complete the one‑click cloud deployment.

After deployment, retrieve the endpoint and token from the service page; refer to the model’s documentation for invocation details.

Use the provided API or the PAI WebUI to interact with the model.

Model Evaluation

Using a video from NVIDIA where a person pours milk into a cup, the model was asked to predict the next plausible action. The model correctly inferred that after pouring, the most reasonable next step is to place the milk bottle back on the countertop, demonstrating its ability to understand motion, object interaction, and temporal reasoning.

Full video, question, and model output are provided for reference.

Upcoming Tools

NVIDIA will also release Cosmos Reason-1 tools, including post‑training scripts (SFT + RL). These enable users to fine‑tune the model on their own data, offering a powerful way to create custom physical AI solutions.

Benchmarks show the tools achieve 1‑2× performance gains over open‑source frameworks on small‑scale clusters.

Contact

For more information or model requests, join the PAI‑Model Gallery user group (DingTalk group 79680024618) or follow the provided links.