NVIDIA Releases Cosmos 3 Open Models for Interactive World Simulation

NVIDIA has expanded its Cosmos world model development platform by releasing four new open-weight checkpoints in the Cosmos 3 model family, aimed at interactive world simulation and multi-sensor driving scenes according to the Hugging Face Blog. All four models are post-trained examples built on Cosmos 3 Nano, feature 16 billion trainable parameters, and are licensed under OpenMDW1.1 for both commercial and non-commercial use.

What happened

According to the Hugging Face Blog, the release includes four distinct checkpoints designed for specific simulation and generation workflows:

• Cosmos3-Nano-Transfer-Auto: A bidirectional model for offline synthetic driving data generation that handles up to 11 camera views, fisheyes, and joint RGB-LiDAR generation using World Scenario Map (WSM) controls.
• Cosmos3-Nano-Sim-Auto: A causally distilled streaming variant of Transfer-Auto that generates successive camera video and LiDAR chunks in a few denoising steps for real-time interaction.
• Cosmos3-Nano-Sim-Transfer: A model that takes spatial controls and text prompts from a physics simulator to render color observations consumed by policies.
• Cosmos3-Nano-Sim-Bimanual: A simulator for two-arm robots (such as YAM and AgiBot embodiments) that predicts environment evolution from robot actions or human joysticks using 59 mixed-action control values.

Under the hood, these models utilize a Mixture-of-Transformers backbone that links text-processing and visual-generation networks through temporal-causal attention. The post-training pipeline spans four stages: bidirectional adaptation, teacher forcing, consistency distillation, and self-forcing to prevent error compounding during long rollouts. Running these architectures requires Linux, an NVIDIA GPU, Docker, and the NVIDIA Container Toolkit, with baseline testing performed on a GB200 system.

Why it matters for small AI teams

For small companies and agent-building teams working on robotics or autonomous driving, open-weight foundational world models drastically lower the barrier to entry for synthetic data generation. Building robust physical agents usually requires massive datasets of edge-case scenarios that are expensive and dangerous to capture in the real world. With Cosmos 3's open weights and documented post-training recipes, small teams can adapt models like Transfer-Auto to their custom sensor rigs or fine-tune Sim-Bimanual for specific robotic arms without training foundational visual-generation networks from scratch. The inclusion of distillation recipes (moving from Transfer-Auto to Sim-Auto) also means teams can optimize models for faster streaming inference, making local closed-loop agent evaluation much more practical on modest hardware setups.

What to do this week

1. Review the technical report and clone the relevant open skills for the checkpoints matching your domain—such as Cosmos3-Nano-Sim-Bimanual for robotic manipulation or Transfer-Auto for custom sensor layouts.
2. Verify your local or cloud infrastructure meets the technical prerequisites, ensuring your development nodes are configured with Linux, Docker, and the NVIDIA Container Toolkit.
3. Set up a pilot data pipeline to evaluate how your existing raw capture formats map to the required control schemas, such as the 59-value mixed-action layout for bimanual embodiments.

Sources

• New NVIDIA Cosmos 3 Open Models for Interactive World Simulation — Hugging Face Blog