NVIDIA Cosmos 3 Super

NVIDIA Cosmos 3 Super is the flagship of NVIDIA's third generation of world foundation models — models built not to answer questions but to predict what happens next in a physical scene. Weights were published on Hugging Face on 31 May 2026 under the OpenMDW 1.1 licence, which permits commercial use; the family had been announced two and a half months earlier, in March 2026, alongside NVIDIA's robotics partner programme. What separates Cosmos 3 from a video generator is the action channel. Alongside text, images, video and audio, the model accepts an action trajectory — a per-frame sequence of joint positions, gripper state or camera pose — and it emits one too. That gives two operations a robotics team actually needs: forward dynamics (here is the scene, here is the movement I intend, show me what happens) and inverse dynamics (here are two frames, tell me what movement produced them). The same checkpoint does both, because both are just modalities in and modalities out. The architecture is a Mixture-of-Transformers with two towers. An autoregressive transformer generates discrete tokens, so text comes out by ordinary next-token decoding. A diffusion transformer handles everything continuous — images, video, audio and actions — by iterative denoising. NVIDIA's claim is that this keeps each modality on the generation mechanism that suits it, rather than forcing everything through one. The published parameter count is 64 billion; the checkpoint on Hugging Face weighs in at 64,615,003,632 parameters in BF16. Action support is not generic — it is listed per embodiment, with a fixed dimensionality for each: a single Franka Panda arm with a RobotiQ gripper is 10D, a dual-arm setup 20D, AgiBot 29D, egocentric human motion 57D, an autonomous vehicle 9D. A robot outside that list cannot simply be plugged in. Two limits are worth reading before treating this as a general video model. Audio input is capped at half a second. Video input is capped at five frames — the model is meant to be shown a situation, not a film. Output is the opposite: 5 to 400 frames, 189 by default, at 256p, 480p or 720p. The reasoning side of the model is more conventional, with a context window of 256,000 tokens and video recommended at 4 fps. The training set is the part that explains the compute bill. NVIDIA reports 767 million image samples and 348 million video samples on the generation side, and 22 million text, 19 million image and 1 million video samples on the reasoning side, collected as a hybrid of sensor, synthetic and automated sources. The family is deliberately tiered rather than versioned: Cosmos3-Nano at 16B and Super at 64B shipped together on 31 May 2026, followed on 20 July 2026 by Cosmos3-Edge at 4B and by four-step distilled variants of the Super image and video generators. Older Cosmos generations — Predict, Transfer and Reason — remain available and are not withdrawn.

▸Videos

No videos yet.

▸Reviews