NVIDIA Cosmos 3 Nano Open-Weight Models Explained: Four Free Checkpoints for Robot and Driving Simulation

Image: NVIDIA (via Hugging Face) — Cosmos3-Nano interactive world simulation (official announcement visual from NVIDIA's 5 November 2026 Hugging Face blog post).

TL;DR — NVIDIA Cosmos 3 Nano open weight models explained in one line: on 5 November 2026 NVIDIA released four free 16-billion-parameter checkpoints that generate interactive driving and robot simulations — multi-camera video plus LiDAR — under an OpenMDW1.1 licence that permits commercial use. They are developer tools, not apps: you run them on your own NVIDIA GPU under Linux.

📌 KEY HIGHLIGHTS

• Four new checkpoints, announced 5 November 2026: Cosmos3-Nano-Transfer-Auto, Sim-Auto, Sim-Transfer and Sim-Bimanual, all post-trained from Cosmos 3 Nano.
• 16 billion trainable parameters each, built on NVIDIA's Cosmos 3 world-model platform for physical AI.
• Free for commercial use: released under the OpenMDW1.1 licence, which permits commercial and non-commercial use under its terms.
• Driving and robots, covered: Transfer-Auto generates synthetic driving data across up to 11 camera views plus LiDAR; Sim-Auto streams it live; Sim-Transfer turns simulator controls into photorealistic video; Sim-Bimanual predicts what a two-arm robot sees next.
• The catch is hardware: Linux, an NVIDIA GPU, Docker and the NVIDIA Container Toolkit are required — there is no cloud demo or consumer app.

What is NVIDIA Cosmos 3 Nano?

Cosmos 3 is NVIDIA's family of "world models" for physical AI — AI systems that understand, simulate and act in the real world, rather than just answering questions or generating pictures. The family launched on 31 May 2026 in three sizes: Edge (4 billion parameters, for on-device robots), Nano (16 billion, for workstation GPUs) and Super (64 billion, for data centres). Cosmos 3 Nano is the middle, practical size: big enough to generate realistic video, small enough to run on a workstation-class GPU such as an RTX PRO 6000.

A world model is different from an ordinary video generator. Instead of turning a text prompt into a clip, it predicts what happens next in a scene when something acts — a car steers, a robot arm moves, a traffic light changes. That makes it useful for training and testing robots and self-driving systems in simulation, where mistakes cost nothing, instead of on real roads and factory floors, where they cost a lot. Until now, that kind of interactive simulation mostly lived inside closed company labs. NVIDIA's move is to publish the weights openly, so any team can download, adapt and build on them.

NVIDIA Cosmos 3 Nano open weight models explained: the four checkpoints

All four checkpoints are post-trained versions of the same Cosmos 3 Nano base, each specialised for one simulation job. NVIDIA published them on 5 November 2026 on its Hugging Face blog, with the weights on Hugging Face, code on GitHub, and a four-stage training recipe so teams can adapt them to their own sensors and robots.

Checkpoint	What it generates	Best for
Transfer-Auto	Offline synthetic driving data: synchronised video across up to 11 camera views, including fisheyes, plus joint RGB–LiDAR, with weather, lighting and time of day varied on demand	Building training datasets for self-driving systems
Sim-Auto	Streaming camera video and LiDAR in small chunks, a few denoising steps at a time, for long interactive driving rollouts	Live, closed-loop driving simulation
Sim-Transfer	Photorealistic RGB video from a physics simulator's spatial controls (depth, segmentation, edges, blur) plus a text prompt	Making simulator output look real enough to train vision policies on
Sim-Bimanual	The next video frames of a scene in response to two-arm robot or camera actions (YAM and AgiBot embodiments)	Testing two-arm manipulation without a physical robot

The important design choice is that appearance and physics are kept separate. In Sim-Transfer, for example, your simulator still computes motion, contacts and physical state; the model only renders what the robot's camera would see. That means the simulation cannot quietly invent physics that flatters your robot — a common failure when a video model is asked to do the whole job itself.

How the streaming generation works

The checkpoints generate video in short chunks: controls plus generated history go in, a short segment comes out, and the result becomes history for the next chunk. Smaller chunks let new controls — a steering input, a gripper movement — take effect more often, which is what makes the simulation interactive rather than a pre-recorded clip. NVIDIA's published performance examples put Sim-Bimanual at 48.1 frames per second at 480p on a single GB300 GPU with 166 milliseconds of latency, Transfer-Auto at 15.85 frames per second on one GB200, and the multi-view driving setup at 16.7 frames per second across 16 GB200 GPUs. Those are NVIDIA's own figures, not independent measurements, so treat them as the maker's claim.

Behind the release is a four-stage training recipe — bidirectional adaptation, teacher forcing, consistency distillation and self-forcing — which NVIDIA also published, so teams can post-train the checkpoints on their own driving logs or robot data instead of starting from scratch. Transfer-Auto is designed as the starting point: adapt it to your camera layout and LiDAR setup, then distil it into the streaming Sim-Auto form. NVIDIA says Transfer-Auto can generate clips of up to 400 frames at 480p or 720p.

Image: NVIDIA — a robot arm in a Cosmos 3 manipulation demo (official visual from NVIDIA's Cosmos 3 technical blog).

Why it matters for India readers

India has a growing robotics and autonomous-systems research base — IIT labs, startups working on warehouse automation and delivery robots, and self-driving research — and simulation is where most of that work starts. The practical value here is synthetic data: Transfer-Auto can generate driving scenes in rain, at night, or in dense mixed traffic without anyone having to collect that footage on real Indian roads, where those conditions are also the hardest and riskiest to record. Because the licence permits commercial use, a small Indian team can build a product on these weights without paying a model licence fee.

Two honest caveats. First, the hardware barrier is real: you need an NVIDIA GPU, Linux, Docker and the NVIDIA Container Toolkit, and NVIDIA's own test setup used a GB200 — data-centre hardware that few Indian students or indie developers have at home. Second, these are research tools with no independent evaluation yet; no one outside NVIDIA has measured how faithfully the generated worlds support real policy learning. If you are a general user, this release changes nothing for you directly — there is no app to download and no chatbot to try.

Should you care?

✅ Verdict: care about it if you build robots, self-driving systems or simulation tooling — this is the most capable open-weight world-simulation release of the year, and it is free to use commercially. Skip it if you are a general user: there is nothing to install or try without a Linux machine and an NVIDIA GPU, and the value is in what developers build on top of it.

Price, availability and how to try it

The checkpoints are free to download from NVIDIA's Cosmos 3 collection on Hugging Face, with code, post-training recipes, notebooks and "open skills" for AI coding agents on GitHub. There is no price, no waitlist and no India-specific restriction — open weights are downloadable worldwide. The system requirements are Linux, an NVIDIA GPU, Docker and the NVIDIA Container Toolkit; NVIDIA's published vLLM-Omni setup was tested on one GB200, which the company says is not a minimum-memory specification. The release follows NVIDIA's early-November update extending the open Cosmos Framework to GR00T, Qwen and pi0.5 robot models, so the four checkpoints land in a toolkit that is widening fast.

FAQs

What is NVIDIA Cosmos 3 Nano?

Cosmos 3 Nano is NVIDIA's 16-billion-parameter open world model for physical AI, released on 31 May 2026. It understands and generates text, image, video, audio and action sequences, and is designed for workstation GPUs such as the RTX PRO 6000, for robotics, autonomous driving and smart-space applications.

What are the four new Cosmos 3 Nano checkpoints?

Announced on 5 November 2026, they are Transfer-Auto (offline synthetic driving data with up to 11 camera views plus LiDAR), Sim-Auto (streaming driving observations), Sim-Transfer (simulator controls plus text into photorealistic video) and Sim-Bimanual (video predicted from two-arm robot actions). All four are post-trained from Cosmos 3 Nano and carry 16 billion trainable parameters.

Is NVIDIA Cosmos 3 free for commercial use?

Yes — the checkpoints are released under the OpenMDW1.1 licence, which NVIDIA says permits commercial and non-commercial use under its terms. Teams considering production deployment should read the actual licence text before relying on that summary.

What hardware do you need to run Cosmos 3 Nano?

Linux, an NVIDIA GPU, Docker and the NVIDIA Container Toolkit. NVIDIA's published vLLM-Omni setup was tested on one GB200, which is not stated as a minimum specification, and the models are optimised for NVIDIA Ampere, Hopper and Blackwell architectures.

Takeaway: NVIDIA has taken interactive world simulation — until now mostly a closed-lab capability — and published it as free, open-weight checkpoints that any robotics or driving team can download and adapt. The models are genuinely useful, the licence is genuinely open, and the only real barrier is the NVIDIA GPU they need to run on.

If you build robots or driving systems, start with Transfer-Auto in NVIDIA's Cosmos 3 collection on Hugging Face and test it on a small set of your own sensor data before committing — the checkpoints, post-training recipes and notebooks are all there.

Sources

• Hugging Face / NVIDIA — "New NVIDIA Cosmos 3 Open Models for Interactive World Simulation" (primary announcement, 5 November 2026)
• NVIDIA Technical Blog — "Develop Physical AI Reasoning, World, and Action Models with NVIDIA Cosmos 3" (primary source, Cosmos 3 family background)
• NVIDIA — Cosmos physical AI development platform (model sizes and platform overview)
• NotaTechGuy — "NVIDIA releases four open-weight Cosmos 3 world models" (release summary, 6 November 2026)
• Extrapolator — "NVIDIA releases four Cosmos 3 open-weight models for world simulation" (checkpoint details, 5 November 2026)

More AI news on TechByCountry: Thinking Machines' Inkling 2 Small, a 594B AI built for coding and autonomous work; Baidu ERNIE 5.5 Flash Base, the 280B open model with a licence catch; Gemini Omni 1.2 Flash, Google's 15-second AI video model at $0.07 a second.