NVIDIA released Cosmos 3 Edge at SIGGRAPH 2026 this week — a 4-billion-parameter open world model for physical AI that runs on a consumer GeForce RTX GPU. The weights are on Hugging Face. Commercial derivatives are allowed. And NVIDIA says developers can adapt it to a specific robot or sensor stack in roughly one day. Most physical AI has demanded cloud connectivity or datacenter hardware; this changes that calculus — with some important caveats worth reading before you plan a timeline around it.
Two Towers, One Very Specific Goal #
Cosmos 3 Edge uses a two-tower Mixture-of-Transformers architecture — not Mixture of Experts, a distinction NVIDIA has been careful to make. The autoregressive tower handles vision and text, giving the model its reasoning and scene-understanding capabilities. The diffusion tower handles generation: video, audio, and critically, native robot control signals — joint angles and trajectories output directly, not parsed from text. Both towers share multimodal attention layers, which is what aligns language, video, audio, and action into a coherent physical-world understanding.
Unlike the Nano (16B) and Super (64B) variants in the Cosmos 3 family, Edge was trained from scratch — not initialized from pre-trained Qwen3-VL weights. That architectural independence is part of why it fits in 4B parameters without giving up meaningful capability at the 4B scale: it currently ranks first on VANTAGE-Bench for vision analytics among same-size models. On a Jetson Thor, it outputs actions at 15 Hz — fast enough for real-time robot control without a cloud round-trip. For context, ByteIota covered the Cosmos + GR00T rollout when NVIDIA first staked its physical AI claim in early 2026; Edge is the edge-deployment chapter of that story.
What Hardware You Actually Need #
For local experimentation, a GeForce RTX 3070 or better is sufficient. The full 4B model runs on a single consumer GPU, which genuinely puts physical AI prototyping within reach of individual developers. If you only want the companion 2B reasoning module — which ships as a separate component — a Jetson Orin 8GB covers it.
For production deployment, NVIDIA is targeting its new Jetson Thor modules: the T2000 at 400 FP4 TFLOPS and the T3000 at 865 FP4 TFLOPS. RTX PRO GPUs and DGX systems work for workstation and server-side inference. The framing NVIDIA uses internally is accurate: your RTX workstation is the development sandbox; Jetson is what goes into the robot.
Getting Started: Three Steps #
The weights are live now via the official Hugging Face introduction post. Setup:
huggingface-cli download nvidia/Cosmos3-Edge
git clone https://github.com/NVIDIA/cosmos
cd cosmos
pip install diffusers accelerate torch torchvision transformers \
cosmos_guardrail huggingface_hub imageio imageio-ffmpeg av
python cookbooks/cosmos3/run_inference.py \
--model nvidia/Cosmos3-Edge \
--no-guardrails \
--torch-backend auto
The guardrails require access to gated safety models on Hugging Face — you will need to accept their license and set your HF_TOKEN
environment variable for full compliance. For early experimentation, --no-guardrails
removes that dependency. Once you are shipping anything externally, enable them. NVIDIA also offers pre-built containers via its developer portal, which is the faster path if you want to skip CUDA version matching. The full inference codebase and cookbooks live at github.com/NVIDIA/cosmos.
Post-Training: What “One Day” Actually Means #
NVIDIA claims developers can post-train Cosmos 3 Edge for a specific robot or sensor stack in about a day. That claim is real, but the hardware context matters: “one day” refers to training time on H100 GPUs or a DGX Station — not on the RTX card you use for inference. If you are planning a production timeline around this, budget for cloud H100 time or a DGX rental, not your workstation.
The practical entry point for most developers is the Cosmos 3 Edge Policy (DROID) checkpoint, available at huggingface.co/nvidia/Cosmos3-Edge-Policy-DROID
. It ships pre-trained on the DROID dataset for robot manipulation — pick-and-place tasks — and comes with post-training scripts you can adapt to your own data. Starting from DROID is faster and cheaper than starting from the base model if your use case is manipulation. NVIDIA’s developer guide walks through the full post-training workflow.
The License: Commercial-Friendly, With a Catch #
Cosmos 3 Edge ships under the OpenMDW-1.1 license, which the Linux Foundation released specifically for AI models. Commercial use and derivatives are permitted. The requirement is attribution: anything you ship that incorporates Cosmos weights must display “Built on NVIDIA Cosmos” somewhere visible — a UI element, about page, documentation, or website satisfies this.
OpenMDW-1.1 is not Apache 2.0 or MIT. It carries acceptable-use restrictions and specific redistribution terms that differ from standard permissive licenses. Before building a commercial product on Cosmos weights, read the actual license file — not a summary of it. NVIDIA has adopted OpenMDW across Cosmos, Isaac GR00T, Ising, and Nemotron, so the investment in understanding it covers multiple ecosystems.
The Timing Question #
The weights and inference code are live and functional today. The full production SDK, including finalized documentation and all toolchain integrations, is expected by end of September 2026. If you are building toward a product launch, September is the more reliable baseline. If you are prototyping or exploring what physical AI can do with a model this size, there is no reason to wait.
The robotics industry is paying attention: FANUC, Yaskawa, and Kawasaki have committed to building on the Cosmos 3 platform, and Fujitsu has collaborative control work underway. For developers, Cosmos 3 Edge is the most accessible entry point yet into physical AI — a 4B model on a GPU you probably already own, with commercial licensing that does not require a legal review before you write your first fine-tuning script. The SIGGRAPH 2026 announcement covers the full scope of what NVIDIA is building across the physical AI stack.