Cosmos 2.5

Cosmos 2.5 refers to two companion model releases — Cosmos Predict 2.5 and Cosmos Transfer 2.5 — which form part of NVIDIA's Cosmos physical AI simulation platform. Both were released at CES 2026 and are part of the broader NVIDIA Physical AI Stack. [NVIDIA AI Models 2026 Guide]

Cosmos is designed to generate realistic synthetic data at scale, bridging the gap between simulated environments and real-world physical AI deployment. It complements GR00T N1.7 (humanoid robotics) and Alpamayo R1 (autonomous driving) by supplying the training data those models need to generalise across diverse conditions. [NVIDIA AI Models 2026 Guide]

Cosmos Predict 2.5 and Cosmos Transfer 2.5 perform the following functions: [NVIDIA AI Models 2026 Guide]

• Synthetic video generation — produces realistic video from a single image
• Multi-camera driving scene creation — generates diverse autonomous-vehicle training scenarios
• Edge-case simulation — synthesises rare or dangerous environments that are costly to capture in the real world
• Physical reasoning — models how objects and agents behave under physical constraints

Together they enable teams to generate large volumes of synthetic training data across diverse conditions, dramatically reducing the need for expensive real-world data collection. [NVIDIA AI Models 2026 Guide]

Cosmos Reason, a separate component of the Cosmos platform, powers AI reasoning for agents in traffic and workplace productivity contexts, and is used by partners such as Salesforce, Milestone, Hitachi, and Uber. [NVIDIA AI Models 2026 Guide]

The following organisations were cited as production users of Cosmos 2.5 at the time of its release: [NVIDIA AI Models 2026 Guide]

Cosmos 2.5 models are available free on Hugging Face, GitHub, and NVIDIA Build (build.nvidia.com). Like all NVIDIA open models in the 2026 portfolio, they are released under licenses — Apache 2.0, MIT, or the NVIDIA Open Model License — that permit commercial use. [NVIDIA AI Models 2026 Guide]

Developers can access Cosmos through NVIDIA NIM (Inference Microservices) for managed cloud deployment without needing to provision their own GPU infrastructure. [NVIDIA AI Models 2026 Guide]

NVIDIA's open-model strategy is deliberate: releasing powerful simulation tools at no cost drives developer adoption onto NVIDIA hardware. As the source notes, NVIDIA is "not just releasing models — they are releasing the entire recipe," including one of the largest open training datasets in AI history. [NVIDIA AI Models 2026 Guide]

Cosmos 2.5 is part of the NVIDIA Open Model Ecosystem and sits at the heart of NVIDIA's long-term thesis: that the next wave of AI value will be created in physical AI — robotics, autonomous vehicles, and industrial systems — rather than pure language modelling. [NVIDIA AI Models 2026 Guide]

For the competitive landscape and how Cosmos compares to alternatives, see NVIDIA vs. OpenAI, Google, AMD & Meta: AI Competitive Landscape 2026. No other company was cited as offering an equivalent open-source physical AI simulation platform at the time of publication. [NVIDIA AI Models 2026 Guide]

Jensen Huang previewed Cosmos 3 at GTC in March 2026, the next generation of the world foundation model platform. It is expected to deliver improved physical reasoning and higher simulation fidelity for both robotics and autonomous vehicles. No release date was confirmed. [NVIDIA AI Models 2026 Guide]