Curiosity · AI Model
NVIDIA Cosmos
Cosmos is NVIDIA's platform of world foundation models (WFMs) aimed at physical AI. It includes diffusion and autoregressive variants (4B to 14B parameters) that ingest image/video prompts plus text or trajectories, and synthesise physically plausible continuations — useful as a pre-trained base for policy learning, synthetic training data, and closed-loop simulation of robots and self-driving cars. Released with open weights and a dedicated tokenizer / guardrails stack at CES 2025.
Model specs
- Vendor
- NVIDIA
- Family
- Cosmos WFM
- Released
- 2025-01
- Context window
- 1 tokens
- Modalities
- text, vision, video
Strengths
- Open weights under NVIDIA's community licence
- Multiple model sizes and two modelling paradigms (diffusion / autoregressive)
- Tight integration with NVIDIA's Omniverse and Isaac toolchain
- Shipped with tokenizer, guardrails, and evaluation suite
Limitations
- Physical realism is still approximate — not a substitute for a rigid-body simulator for safety-critical loops
- Inference is GPU-hungry — H100-class hardware for reasonable latency
- Evaluation on long horizons still degrades in fidelity
- Not a ready-to-deploy policy — it's a pre-training / data resource
Use cases
- Synthetic data generation for robot and AV training
- Closed-loop simulation with physics-aware dynamics
- Pre-training backbone for VLA and driving policies
- Scenario augmentation for safety testing
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| Video fidelity (internal, 14B diffusion) | state-of-the-art at release | 2025-01 |
| Downstream policy transfer (Cosmos pre-train vs scratch) | meaningful sample-efficiency gains | 2025-01 |
Frequently asked questions
What is NVIDIA Cosmos?
Cosmos is NVIDIA's family of world foundation models designed to generate physically plausible video futures, primarily used as pre-training, synthetic data, and simulation backbones for physical AI applications.
How does Cosmos relate to Omniverse and Isaac?
Omniverse/Isaac provide deterministic physics simulation; Cosmos provides learned world models that generate diverse, photoreal futures. NVIDIA positions them as complementary tools.
Is Cosmos open source?
Weights are released under NVIDIA's open model licence with safety/commercial restrictions, and the tokenizer and eval stack are open too.
Sources
- NVIDIA Cosmos landing page — accessed 2026-04-20
- Cosmos technical report (arXiv) — accessed 2026-04-20