// the find
NVIDIA/cosmos
NVIDIA Cosmos is an open platform of world models, datasets, and tools that enables developers to build Physical AI for robots, autonomous vehicles, smart infrastructure, and more.
NVIDIA Cosmos is a family of omnimodal world models for physical AI — one architecture that either reasons over video and text (the Reasoner) or generates video, sound, and action sequences (the Generator), aimed at robotics, autonomous driving, and simulation research. This repo is the inference and evaluation hub; training and data pipelines are pushed out to separate NVIDIA repos (Cosmos Framework, Cosmos Curator).
The project is honestly scoped — this repo is notebooks, model cards, and eval suites, with fine-tuning explicitly punted to Cosmos Framework rather than crammed in here. It ships three real size tiers (64B/16B/4B) mapped to actual hardware, from H200/B200 down to Jetson Orin, instead of a single flagship checkpoint nobody can run. Backend support is broad and current — Diffusers, vLLM, vLLM-Omni, TensorRT-LLM, SGLang, and NIM are all wired up, so you're not locked to one serving stack. Benchmark suites like Physics-IQ, PAIBench, and VLMEvalKit are linked directly rather than just asserted in a table.
Licensing is NVIDIA's own OpenMDW-1.1, not a standard OSI license, and the required Guardrail model is gated behind a Hugging Face access request plus manual approval before the quickstart even runs. The hardware floor is steep — even the smallest 'Edge' tier targets Jetson AGX Orin/Thor, so there's no path to trying this on a consumer GPU or CPU. The repo itself holds almost no source code; it's mostly notebooks plus sample LeRobot datasets (parquet and mp4 files) checked directly into the tree, making for a heavy clone for something billed as a cookbook. The quickstart also silently pulls a 16B checkpoint on first run with no size or time warning beyond a passing comment.