NVIDIA launches Cosmos 3, an open frontier omnimodel for physical AI — with a creative-tools coalition behind it

NVIDIA released Cosmos 3 on May 31, 2026 at GTC Taipei — an open frontier omnimodel for physical AI built on a mixture-of-transformers architecture that combines vision reasoning and multimodal generation in one model. Cosmos 3 generates text, images, video, ambient sound, and action trajectories together — what NVIDIA calls "the world's first fully open omnimodel" — and ships with open weights available on Hugging Face, deployable as NVIDIA NIM microservices or via build.nvidia.com. For a field where frontier video generation has largely stayed closed and proprietary, the open licensing is the structural signal: NVIDIA is positioning Cosmos 3 as infrastructure, not a gated API product.

Three variants, two available now. Cosmos 3 ships in three tiers. Cosmos 3 Super is the highest-accuracy variant — optimised for post-training robotics and autonomous vehicle models where physics fidelity matters most. Cosmos 3 Nano is the fast tier — high-quality video and action reasoning in fractions of a second, designed for real-time use cases. Cosmos 3 Edge is coming soon, targeting real-time inference at the edge for on-device deployment in robots and autonomous systems. Super and Nano are available now; Edge has no confirmed date.

What the model actually does. Cosmos 3 operates across five modalities simultaneously: text, images, video, ambient sound, and action. The action outputs are numerical — joint angles, gripper positions, trajectory points — which makes the model usable as the reasoning and generation layer for robotic policy development, not just video creation. This is what separates Cosmos 3 from commercial video tools: it can generate a video of a manipulation task and simultaneously output the action data that would make a physical robot execute the same motion. NVIDIA benchmarks Cosmos 3 first among open models on Physics-IQ, PAI-Bench, and R-Bench for world generation; first on RoboLab and RoboArena for action policy; and first on VANTAGE-Bench and TAR for vision understanding. Jensen Huang's framing at the launch: "The Cosmos 3 family of open, frontier omnimodels gives developers a generational leap in ability to build robots, autonomous vehicles and vision AI."

The Cosmos Coalition — and why creative-AI operators should pay attention. Alongside the model launch, NVIDIA announced the Cosmos Coalition: a founding group of companies building on top of Cosmos 3. The robotics members — Agile Robots and Skild AI — are expected. The media and creative-AI members are more relevant for ToolNav operators: Black Forest Labs (the team behind the FLUX image generation models), LTX (AI video generation), Runway (which shipped its own GWM-1 world model in December 2025 and now carries a $5.3B valuation), and Generalist. These are tools with direct overlap with AI content creation workflows. The Coalition's likely path is that these companies build world-model-powered features on top of Cosmos 3 — improved physics accuracy in generated video, better motion coherence, synthetic scene generation for training. For operators using FLUX or LTX-based tools today, Cosmos 3 is the upstream infrastructure that may power the next generation of those capabilities.

What Cosmos 3 is not, yet. Cosmos 3 is not a turnkey creative tool. There is no Cosmos 3 interface comparable to Midjourney or InVideo. Accessing it requires Hugging Face, a NIM microservice, or a supported infrastructure provider (Baseten, CoreWeave, Microsoft Azure, Nebius, Deep Infra, Classmethod). The creative-AI impact will arrive through the Coalition members' products — not through a direct Cosmos 3 consumer experience. The benchmarks are for open models; how Cosmos 3 compares to Sora, Veo, or Kling at the product level is not established by NVIDIA's current benchmark set.

Where it sits in the landscape. For the AI video and image generation category, Cosmos 3's significance is indirect but real. The tools your readers pay for are built on top of foundation models like this one. NVIDIA going open at the world-model layer sets a capability baseline that the entire category is now building toward. The Coalition is the mechanism through which Cosmos 3 shows up in the products creators actually use.