NVIDIA's 4B Cosmos 3 Edge video model runs on a Mac

NVIDIA's 4-billion-parameter Cosmos 3 Edge world model now runs on Apple silicon. An unofficial MLX port of its video generator appeared on Hugging Face on 7 November, published by NetProjects, which states it is not affiliated with NVIDIA.

The port is the new part. NVIDIA's own model card dates the Edge checkpoint's release to 20 July 2026, and the Cosmos 3 GitHub landing page lists Edge under July 2026 as the 4B tier for on-device deployment.

NetProjects' 8-bit build is 5.0 GB, with the transformer's linear layers quantized to 8-bit and the VAE kept in BF16; a BF16 build is 7.6 GB. Measured on a MacBook Pro M5 Max with 128 GB of memory on 7 November, NVIDIA's full image-to-video example runs end to end in 113 seconds with an 18.8 GB peak, at 832x480, 121 frames, 24 fps and 20 steps. The port covers the video generator only: text, or text plus a first frame, in, and video or a still image out.

Cosmos 3 Edge is the smallest of NVIDIA's three Cosmos 3 world models, against the 16B Nano and the 64B Super. Its card describes an omnimodal model on a Mixture-of-Transformers architecture, built for Physical AI: robotics, autonomous vehicles and smart spaces. It ships under the OpenMDW 1.1 license, which the card says permits commercial and non-commercial use.

The generator takes text up to 4,096 tokens, images, and action trajectories. Video comes out at 256p or 480p, 12 to 30 fps and 50 to 150 frames. NVIDIA's August update sets the image-to-video recipe at 832x480, 121 frames, 20 denoising steps, guidance 6.0 and flow shift 12.0. The card lists support for NVIDIA Ampere, Blackwell and Hopper chips, says Linux is the only operating system it has tested, and that only BF16 precision is tested.

On speed, that same update reports Edge generating image-to-video at 6.71 frames a second against the Nano's 2.66, both at 480p and 24 fps on one H100, with a lower PAIBench-G score (81.1 against 82.4). NVIDIA's release post frames Edge as a world action model that works at robot-control resolution (640x360) and generates 32 actions per inference on Jetson Thor at 15 Hz.

That scope is the caveat for a creator. NVIDIA built the model for physically coherent motion in robot and driving scenes, not for arbitrary cinematic prompts, and the card warns of temporal inconsistency, unstable camera or object motion and action-state drift in long, high-resolution outputs. It has no explicit physics simulator, so contact and object permanence are only approximated.

The port and both its builds are on Hugging Face under netprojects, and NVIDIA's own checkpoint is on the nvidia/Cosmos3-Edge page.