Local Ai
Post-Train NVIDIA Cosmos 3 Edge for On-Device Robot Control
NVIDIA’s Cosmos 3 Edge is a 4‑billion‑parameter omni‑model (with a 2‑billion‑parameter Nemotron reasoner) that can run on an NVIDIA Jetson Thor, providing on‑device policy inference for robot manipula
NVIDIA’s Cosmos 3 Edge is a 4‑billion‑parameter omni‑model (with a 2‑billion‑parameter Nemotron reasoner) that can run on an NVIDIA Jetson Thor, providing on‑device policy inference for robot manipulation. The model, pre‑trained on the same physical‑world datasets as Cosmos 3 Nano and Super to capture fundamental object dynamics, is post‑trained within the cosmos‑framework repo to generate closed‑loop control actions via a receding‑horizon loop. A tutorial demonstrates how to serve the trained policy on Jetson Thor, run inference, and evaluate performance in simulation, with checkpoints hosted on HuggingFace.
Related
- Model Quantization: Post-Training Quantization Using NVIDIA Model Optimizer
- GELATO: Generative Entropy- and Lyapunov-based Adaptive Token Offloading for Device-Edge Speculative LLM Inference
- The Cosmos omnimodel family of models - 3 variants Edge(4B) , Nano(16B) , Super (64B)
Source: NVIDIA Developer | 2026-08-19