NVIDIA Cosmos 3 — open-source world model for embodied AI
NVIDIA is betting that the future of robotics lives at the edge, not in the cloud. At the RSS 2026 conference in Sydney, the company's physical AI team laid out detailed technical specs for Cosmos 3 — a fully open-source world foundation model that unifies text, vision, audio, and action in a single Transformer architecture.
Cosmos 3 replaces the fragmented patchwork of separate vision, language, and control models with one unified system. Max Li, NVIDIA's physical AI lead, compared the old approach to assembling a PC from mismatched parts — developers had to stitch together VLMs, LLMs, and control algorithms manually. Cosmos 3 does it all in one MoE model: it understands the world (like a VLM), simulates future outcomes (generating video and audio), and executes actions (producing motor commands). The model ships in three sizes: a 4B-parameter Edge variant that runs in real time on Jetson devices without a server GPU, a 16B Nano, and a 64B Super for research and simulation. Trained on roughly 22 million samples for the reasoner and 2 million for the generator — equivalent to about 10 million hours of video — the team also added inverse dynamics: given a video of robot behavior, Cosmos 3 can reverse-engineer the motor commands that produced it, turning unlabeled video into training data. The model is fully open-source with code, weights, and paper all public. NVIDIA's play is clear: define the standard "brain and body" protocol for embodied AI, the way Android unified mobile, and let the open-source ecosystem fill in the rest. The Edge variant is the real signal here — it means developers can run world-model inference on a $200 Jetson board instead of renting a server rack, which could unlock real-time robotic applications that were previously too expensive to prototype.
What to watch: Cosmos 3 Edge is already advancing Japan's physical AI ecosystem, per The Fast Mode. If developers adopt it the way they did earlier Cosmos versions, NVIDIA could become the default platform for the robotics era — but Chinese competitors are closing fast, with 36Kr reporting one lab overtook Cosmos 3's benchmarks within a day of launch.
Do you think a single open-source world model can unify the fragmented robotics stack, or will specialized models win out? Tell us in the comments.
Sources: Leiphone — NVIDIA Cosmos 3 deep dive at RSS 2026 · The Fast Mode — NVIDIA Cosmos 3 Edge advances Japan's physical AI · Axios — Nvidia's new world model helps robots navigate the world
