NVIDIA Cosmos 3 Edge: A 4B Open World Model Hits Jetson Thor at 15 Hz

by Persephone

NVIDIA's 4B-parameter Cosmos 3 Edge world model runs at 15 Hz on Jetson Thor, generating 32 actions per inference for on-device robot control. Two-tower Mixture-of-Transformers, 2B dense reasoner, Qwen3-VL-compatible conventions, and the third and smallest member of the Cosmos 3 family.

NVIDIA Cosmos 3 Edge: A 4B Open World Model Hits Jetson Thor at 15 Hz

July 21, 2026

NVIDIA released Cosmos 3 Edge yesterday: a 4B-parameter open world model that runs at 15 Hz on Jetson Thor and emits 32 action tokens per inference for on-device robot control (Hugging Face / NVIDIA blog, 2026-07-20). Edge is the third tier of the Cosmos 3 family, sitting one rung below Nano (16B) and Super (64B), and the first designed to run on edge silicon without cloud routing.

Architecture: Two-Tower Mixture-of-Transformers

Cosmos 3 Edge is a Mixture-of-Transformers with two specialized towers sharing multimodal attention layers. The autoregressive tower handles vision-and-text reasoning: it predicts the next token given a video frame, an instruction, and the model’s prior beliefs about the scene. The diffusion tower handles vision-audio-and-action generation: it denoises action trajectories conditioned on the autoregressive tower’s outputs (MarkTechPost, 2026-07-21).

Two concrete details make the design actionable. Edge uses a 2B dense transformer for its reasoner — no MoE routing inside the autoregressive path, which simplifies on-device latency. And the model follows Qwen3-VL-compatible message conventions, so existing Qwen3-VL pipelines, prompt templates, and tool-call shapes slot in without a custom shim (MarkTechPost, 2026-07-21).

The full Cosmos 3 family — Nano, Super, Edge — was introduced at GTC Taipei on May 31, 2026, with Edge as roughly 1/16th the parameter count of Super (MarkTechPost, 2026-06-03). Edge is the first family member targeting edge devices.

Jetson Thor at 15 Hz: What 32 Actions per Inference Buys You

The headline number is 15 Hz with 32 actions emitted per inference. At 15 Hz the model produces a new chunk of 32 action tokens every ~67 ms, which means a single inference step covers roughly 2.1 seconds of robot motion at a 15 Hz control rate. That is long enough for pick-and-place, point-to-point navigation, and the kind of smooth-arm motion that picks up and puts down objects without jerky per-frame replanning (Hugging Face / NVIDIA blog, 2026-07-20).

The throughput lands at a real-time control cadence on Jetson Thor without routing through the cloud. NVIDIA ships a paired policy model — Cosmos 3 Edge Policy (DROID) — fine-tuned for robot pick-and-place; the fine-tunable variant runs post-training on H100 or DGX clusters, and NVIDIA cites roughly one day of post-training on a DGX Station before the policy is deployable on Thor (explainx.ai, SIGGRAPH 2026 analysis).

Benchmarks: #1 on VANTAGE-Bench for Vision Analytics

Cosmos 3 Edge ranks #1 on VANTAGE-Bench for vision analytics among 4B-parameter models, and is the state-of-the-art for robot policy learning in its size class (Cryptonomist, 2026-07-21). VANTAGE-Bench is the public vision-analytics benchmark the field has converged on for measuring world models; a 4B model taking the top spot matters because it is roughly 1/16th the parameter count of Cosmos 3 Super and still beats larger open-world models on the test set.

The implication is not “4B > 64B.” It is that for the specific workload VANTAGE measures — visual reasoning with embodied-action conditioning — the Edge architecture gets most of the way to Super quality at a fraction of the inference cost.

Targets and the New Jetson Modules

Edge targets four hardware tiers: RTX PRO for workstation-class robotics, DGX for training and large-batch offline inference, GeForce RTX for the developer bench, and the Jetson T2000 and Jetson T3000 for in-robot deployment. The T2000 and T3000 are the new Jetson modules NVIDIA announced alongside the model on July 15, 2026 (Techzine, 2026-07-15).

Seeed Studio confirmed launch-day support for both modules the same week, which means carrier boards and integration kits are already shipping rather than queued behind a module-launch backlog (Seeed Studio, 2026-07-15). The Cosmos Coalition — NVIDIA’s partner program for physical-AI deployment — also expanded to Japan in the same announcement.

What This Is Not

Edge is not a replacement for Super or Nano on workloads where they win. Super is the higher-quality choice for offline simulation, large-batch reasoning, and policy-training data generation. Nano is the middle ground for workstation-class robots with discrete GPU access. Edge is for the case where the robot itself is the inference target — no datacenter round-trip, no edge server in the loop.

Edge is also not a chat model. There is no “ask the world model a question” surface. The action tokens are emitted as a trajectory conditioned on observations; the model’s output is motor commands, not prose.

What You Can Do With It Today

Three paths are open right now.

1. Pull the weights. The model is on Hugging Face at nvidia/Cosmos3-Edge under an open license. The DROID policy checkpoint ships alongside. PyTorch and TensorRT inference paths are both supported.

2. Fine-tune on your own robot data. NVIDIA’s reference path is roughly one day of post-training on a DGX Station to specialize the policy for a specific robot platform or task. Smaller fine-tunes on H100 are also documented.

3. Buy the silicon. Jetson T2000 and T3000 modules are orderable through Seeed Studio and the usual NVIDIA partners. Cosmos 3 Edge at 15 Hz is the workload the modules were designed against.

The practical read for robotics teams: a 4B-parameter open world model that runs on-device at 15 Hz is the threshold where closed-loop physical reasoning stops requiring a server rack in the robot’s closet. Cosmos 3 Edge lands exactly on that threshold, with the Jetson T2000 / T3000 silicon to back it and a day-long fine-tuning path for new platforms. The model card and benchmarks are public; the main remaining unknown is how the 15 Hz number holds up on real-world robot platforms with non-ideal camera input and partial occlusions, which is what the next month of integrator deployments will show.

Sources

Prior TopClanker Coverage