DATASHEET // COSMOS
NVIDIA Cosmos
World-foundation models for physical AI.
OPEN SOURCE24–80 GB VRAMRuns locally
Actually FreeWatermark-FreeAPI
Visit NVIDIA CosmosUPDATED 2026-05-12 · DIRECT LINK
HARDWARE REQUIREMENTS //
Runs locally · Datacenter GPU (80 GB+)
Min VRAM
24 GB
Rec. VRAM
80 GB
Min RAM
32 GB
Rec. RAM
64 GB
Disk
80 GB
GPU class
Datacenter GPU
12.0+No Apple SiliconGPU RequiredQuant: FP16, BF16
4B variant on 24 GB; 14B + tokenizer pipeline needs H100-class.
[ EDITORIAL PICK ]
Why we recommend NVIDIA Cosmos
- Genuinely free
Has a free tier you can actually finish a project on, not the 3-credits-then-paywall pattern.
- Top-tier pick
Power-user score 92/100 — consistently rated highly by people who use this every day, not just benchmark chasers.
- 2 quant formats
Supports FP16, BF16 — you can dial VRAM use up or down to match your card.
- Hosted API too
Both self-hostable and available as a hosted API — prototype on someone else's GPU, deploy on yours.
[ EVIDENCE NOTE ]
Documentation-led datasheet
This page summarizes upstream documentation, release information, and editorially reviewed catalogue fields. It is not presented as a hands-on benchmark. Verify changing requirements at the official project; report stale data through our corrections channel.
AT-A-GLANCE SIGNALS //
DERIVED FROM THIS PAGE'S DATA- Install difficultyStandard
A standard local install — download, install dependencies, point at your GPU.
- Hardware comfortEnthusiast
Needs 24 GB minimum — RTX 3090 / 4090 territory.
- EcosystemAPI-first
Exposes a stable API — you can build on top of it programmatically.
- VerificationRecent
Catalogue entry last updated 65 days ago — re-verification due soon.
[ MORE IN THIS NICHE ]
Other heavy video generation tools we rate
Three picks across different tradeoffs — so you don't end up with three near-clones of NVIDIA Cosmos.
What is NVIDIA Cosmos?
Cosmos is NVIDIA's open foundation-model family for "physical AI" — robotics, autonomous vehicles, simulation. Includes diffusion and autoregressive video generators (Cosmos-Predict), tokenizers, and guardrails. Aimed at synthetic-data generation for physical-world ML rather than artistic video.
Pros & cons
✓ PROS
- Purpose-built for synthetic data in robotics / AV pipelines
- Physics-aware video generation (camera motion, object permanence)
- Full pipeline: tokenizer, generator, guardrails
- Open weights at multiple sizes (4B → 14B)
– CONS
- Hardware floor is high — 48 GB+ for the larger variants
- Not aimed at general creative video — different model than Wan/Hunyuan
- Licence is open but not OSI-approved
What's actually free?
NVIDIA Open Model License — permissive but not OSI-OSS.
✓ Actually FreeWatermark-Free
Alternatives
LTX-Video
Real-time-ish open video diffusion from Lightricks.
OPEN SOURCE12–16 GB VRAM
VRAM fit12–16 GB