
Core Announcement and Physical AI Scope
NVIDIA has published and updated nvidia/Cosmos3-Edge on Hugging Face, positioning the 4-billion-parameter release as a compact entry point in its Cosmos 3 omnimodal world foundation model family. According to the official model card, the Cosmos 3 platform was originally released on July 20, 2026 across Hugging Face and GitHub, followed by an August 25, 2026 update that refreshed the Cosmos3-Edge generator checkpoint, runtime defaults, usage examples, and benchmark documentation, with the repository listing carrying a November 12, 2026 timestamp. Released under the OpenMDW1.1 license for both commercial and non-commercial use worldwide, Cosmos3-Edge targets physical AI workflows spanning robotics, autonomous vehicles, and industrial or factory-scale smart spaces. Rather than acting solely as a video generator or a vision-language classifier, the model combines multimodal understanding, future-state prediction, action reasoning, and world simulation inside a single 4B-parameter checkpoint.
Mixture-of-Transformers Architecture and Model Specifications
Technically, Cosmos3-Edge uses a Mixture-of-Transformers (MoT) design built on NVIDIA's Cosmos Framework. The architecture pairs two complementary transformer towers: an autoregressive transformer that produces discrete text tokens via next-token decoding, and a diffusion transformer that synthesizes continuous modalities, including images, video, audio, and robot action trajectories, through iterative denoising. NVIDIA reports that the model was trained on 1.3 billion data points drawn from 393 dataset entries, with 100 percent of that corpus allocated to training while evaluation benchmarks are maintained separately. In its generator configuration, Cosmos3-Edge accepts text strings up to 4,096 tokens, 8-bit sRGB color images at 256p or 480p resolution across five aspect ratios (16:9, 4:3, 1:1, 3:4, and 9:16), and JSON action trajectories formatted as two-dimensional arrays of 16 to 400 frames. Outputs span JPEG stills, MP4 video clips of 50 to 150 frames at 12 to 30 frames per second, JSON action arrays, and text. Separately, the model's reasoner mode expands the text context window to 256K tokens and ingests text alongside images or MP4 video recommended at 4 frames per second, outputting structured chain-of-thought reasoning, 2D and 3D point localization, and bounding-box coordinates with a recommended default of 4,096 or more output tokens.
Embodied Robotics and Developer Workflow Integration
For robotics and simulation teams, the most consequential detail in the release is its explicit embodiment schema. Instead of emitting generic control signals, Cosmos3-Edge standardizes action inputs and outputs as (T, D) arrays mapped to ten named physical and virtual platforms: general camera motion (9D), autonomous vehicles (9D), egocentric motion (57D), single Franka Panda arms with RobotiQ grippers (10D), dual Franka Panda arms with RobotiQ grippers (20D), Agibot (29D), UR robots (10D), Google robot embodiments (10D), WidowX 250 (10D), and UMI (10D). This schema allows developers to condition future visual rollouts on robot joint states, gripper status, or camera poses, or conversely predict candidate action sequences from visual context. On the software side, NVIDIA lists support for vLLM-Omni, vLLM, PyTorch, and Hugging Face Diffusers, enabling teams to plug the 4B weights into existing serving and diffusion pipelines rather than building custom inference runtimes from scratch. However, the Hugging Face model page indicates that ready-to-run code snippets for the standalone Cosmos library are not yet populated, directing developers to the underlying repository files and documentation, while teams upgrading from earlier snapshots are advised by NVIDIA to pull a fresh copy of the August 25 generator checkpoint to preserve reproducibility.
Hardware Requirements, Precision Limits, and Deployment Caveats
Despite its Edge branding and relatively modest 4B parameter count, Cosmos3-Edge carries strict compute and precision boundaries that engineers must account for before targeting field hardware. NVIDIA states that the model is optimized for NVIDIA GPU-accelerated Linux systems and officially lists compatibility only for NVIDIA Ampere, Blackwell, and Hopper microarchitectures; other operating systems have not been tested. Crucially, only BF16 precision is tested and officially supported, while quantized or lower-precision formats such as FP4, FP8, and FP16 remain unsupported. That restriction narrows immediate deployment options on memory-constrained embedded boards that rely on FP8 or INT4 quantization for real-time throughput. Input constraints also apply: grayscale images are rejected in favor of 8-bit sRGB color, and generator video output tops out at 480p and 150 frames per call, positioning the model for short-horizon rollouts and planning rather than high-definition simulation. Finally, while NVIDIA's model card references reasoning and generation benchmark sections and notes an August 25 update to benchmark results, complete quantitative metrics are not presented in the available model-card overview and have not been independently verified here. NVIDIA itself emphasizes that integrating Cosmos3-Edge into production robotics or autonomous vehicles requires additional domain-specific datasets and iterative V-model unit and system-level validation before real-world deployment.
Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-19.



