NVIDIA Cosmos 3 Edge

NVIDIA Cosmos 3 Edge is the small tier of NVIDIA's third generation of world foundation models, published on 20 July 2026, seven weeks after the 16-billion Nano and 64-billion Super. It has 4 billion parameters — 3,858,999,728 in BF16 — and the same Mixture-of-Transformers design as the flagship: an autoregressive tower for text, a diffusion tower for images, video and robot action trajectories. The cuts are not only in size. Cosmos 3 Super accepts text, images, video and audio and can emit audio too; Edge accepts text, images and action trajectories only, and returns text, images, video and actions. The audio channel is gone, and so is video input — Edge is shown a still frame and an intended movement, not a clip. The reasoning side keeps the full 256,000-token context window. What makes this entry worth reading is that NVIDIA published the timings, and they are unusually candid for a model marketed as edge-ready. All figures are for a single GPU at batch size 1, 480p, generating 189 frames. On an H100 SXM the model produces a video from an image in 27.64 seconds, answers a forward-dynamics query — what happens if the arm makes this move — in 3.91 seconds, an inverse-dynamics query in 5.60 seconds, and runs the DROID manipulation policy in 1.41 seconds. On a B200 the policy drops to 0.99 seconds. Move the same model onto hardware a robot can actually carry and the picture changes. On a Jetson AGX Thor T5000 with 128 GB in MAXN mode, image-to-video takes 137.50 seconds, forward dynamics 6.05 seconds, inverse dynamics 7.19 seconds and the DROID policy 6.32 seconds. On a 32 GB Jetson T3000 the video job takes 194.76 seconds and the policy 8.63. That is the honest reading of "edge": this is a model for planning, data generation and offline reasoning aboard a machine, not a controller closing a loop at servo rates. Edge shipped in a group of four checkpoints on 20 July 2026: Cosmos3-Edge itself, Cosmos3-Edge-Policy-DROID for manipulation on the DROID platform, and four-step distilled versions of the Super image and video generators, which use Improved Distribution Matching Distillation to cut generation to four denoising steps. In practice it is the most downloaded model of the family: roughly 894,000 downloads in the 30 days to 26 August 2026, against about 125,000 for Super — a gap that says more about how many people can run a 4-billion model than about which one is better. The licence is OpenMDW 1.1, the same as the rest of the family, and permits commercial use.

▸Videos

No videos yet.

▸Reviews