What Is Cosmos 3 Edge? The 4B World Model That Runs Inside the Robot

Cosmos 3 Edge is a 4-billion-parameter open world model from NVIDIA, built on Nemotron, that runs vision reasoning and robot action generation directly on Jetson edge devices — no cloud round-trip. It launched on July 15, 2026 in Tokyo as the third and final tier of the Cosmos 3 family, joining the 16B Nano and 64B Super that shipped May 31, 2026. Its job is narrow and physical: help a robot see its surroundings, reason about them in real time, and emit joint-angle or trajectory commands locally.

Key Takeaways

• Cosmos 3 Edge is a 4-billion-parameter world model built on NVIDIA Nemotron, launched July 15, 2026 at a Jensen Huang event in Japan (NVIDIA Newsroom).
• It is the on-device tier of the Cosmos 3 family, following the 16B Nano and 64B Super released May 31, 2026 at GTC Taipei.
• Target hardware includes NVIDIA RTX GPUs, DGX systems, and the new Jetson Thor T2000 and T3000 modules (400 and 865 FP4 TFLOPS respectively).
• Outputs are multimodal: text, images, video, ambient sound, and native robot control signals such as joint angles and trajectories.
• Developers can post-train it for a specific robot, vehicle, or sensor stack in about one day using the open Cosmos framework.
• Cosmos 3 ships under the Linux Foundation OpenMDW-1.1 license, which allows commercial derivatives with attribution.
• Cosmos 3 Nano and Super weights are on Hugging Face; the Edge Hugging Face publication date was not confirmed at the July 15 announcement, though the model is available through the open Cosmos framework.

What Is Cosmos 3 Edge?

Cosmos 3 Edge is the edge-deployment variant of NVIDIA's Cosmos 3 physical AI family. It targets the "chip inside the robot's head" — the compute-constrained module that has to perceive, reason, and act without waiting on a datacenter.

At launch on July 15, 2026, NVIDIA described the model as a 4-billion-parameter member of the Cosmos world model family, built on the Nemotron foundation and designed to help "embodied systems see, reason in real time and predict robot actions locally" (NVIDIA Newsroom). The base Cosmos 3 family debuted May 31, 2026 at GTC Taipei with two tiers — Cosmos 3 Nano (16B) and Cosmos 3 Super (64B) — with Edge listed at that time as "coming soon."

Cosmos 3 Edge is now the third tier, and it ships with a notable architectural detail: unlike Nano and Super, which were initialized from pretrained Qwen3-VL weights, Edge is trained from scratch. It is not a chatbot, not an image generator, and not a general foundation model in the LLM sense. It is a World Model whose outputs include direct joint-angle commands for actuators, which makes it a candidate policy backbone rather than a perception module bolted onto something else.

Cosmos 3 Edge at a Glance

Attribute	Value
Developer	NVIDIA
Type	Open world foundation model (physical AI)
Family	Cosmos 3 (Edge / Nano / Super)
Parameters	4 billion
Base	NVIDIA Nemotron (trained from scratch, not initialized from Qwen3-VL)
Architecture	Mixture-of-Transformers (Reasoning + Generation)
Modality (in)	Text, RGB video, depth, point clouds, IMU / motion sequences
Modality (out)	Text, video frames, ambient sound, joint angles, trajectories
Target hardware	NVIDIA Jetson Thor (T2000, T3000), RTX GPUs, DGX systems
Context window	Not disclosed for Edge
License	OpenMDW-1.1 (Linux Foundation)
Pricing	No per-call price (open model; hardware-based)
Availability	Launched July 15, 2026; usable via the open Cosmos framework. Edge-specific Hugging Face publication date not confirmed
Adaptation time	~1 day for a specific robot / sensor / environment

How Cosmos 3 Edge Works

Cosmos 3 Edge inherits the base family's Mixture-of-Transformers design: a Reasoning Transformer pairs with a Generation Transformer inside a single model. The reasoning half understands object interactions, motion, and spatial-temporal relationships; the generation half then produces the next video frames, ambient sound, or, crucially, an action trajectory.

That "think first, then act" ordering is the point of the world-model framing. As NVIDIA's press description puts it, the architecture is meant to "understand object interactions, motion and spatial-temporal relationships before generating video and action trajectories" (NVIDIA Newsroom). An LLM predicts the next token; a Cosmos 3 model predicts the next state of a physical scene.

What Edge changes is not the architecture but the footprint. At 4B parameters it is roughly a quarter the size of Nano and 1/16th the size of Super, small enough to fit on the Jetson Thor modules NVIDIA also unveiled that week: the T3000 (865 FP4 TFLOPS, 32 GB LPDDR5X, 8-core Arm CPU) and the entry-level T2000 (400 FP4 TFLOPS, 16 GB), per NVIDIA's Jetson Thor announcement (blogs.nvidia.com). Emulation is available now through JetPack; the physical modules ship in Q1 2027.

The one durable claim worth quoting: Cosmos 3 Edge is the first tier of the Cosmos 3 family whose output pipeline is meant to complete inside the robot's own compute envelope, not a datacenter round-trip.

What Makes It Different

Three properties separate Cosmos 3 Edge from the physical AI stacks it targets:

• Action as a first-class output. Cosmos 3 does not treat motion as pixel deltas in a generated video; it emits native joint angles, gripper positions, and trajectory points. Edge inherits this and runs it locally.
• One-day adaptation. NVIDIA states developers can post-train the model for a specific robot, vehicle, sensor set, or environment "in about a day" using the open Cosmos framework — a claim rooted in the model's small size and the availability of Omniverse-generated simulation data.
• Open license with commercial rights. The OpenMDW-1.1 license permits commercial use and derivatives with a "Built on NVIDIA Cosmos" attribution, which is a step further than weight-only releases.

What You Can Do With Cosmos 3 Edge

The concrete use cases NVIDIA and coalition members have named cluster around three workloads:

1. On-device perception and reasoning for robots and vision AI agents — understanding a scene fast enough to act on it without cloud latency.
2. Robot policy models. Post-trained Cosmos 3 Edge instances can serve as the local action policy for a specific arm, mobile base, or humanoid platform.
3. Vision AI agents built with the new NVIDIA Metropolis libraries and skills that NVIDIA announced alongside Edge, which the company claims accelerate agentic vision development at least 6x.

Early community reads emphasize latency as the differentiator. One observer, writing after the Tokyo announcement, framed the significance as "on-device Jetson Thor world model for vision/robot control (no cloud latency)" — a positioning consistent with the Fujitsu-led collaborative control work that FANUC, Yaskawa, and Kawasaki intend to build on top of the platform.

If you are prototyping robot-facing generative pipelines and want a hosted video-generation option to compare against Cosmos-style world simulation, an image-to-video model like Seedance 2.5 is a common baseline for the video-generation half of the workload — Cosmos 3 Edge itself is not offered as a hosted API.

How Cosmos 3 Edge Compares

Within the Cosmos 3 family, tiering is by hardware target:

Tier	Parameters	Target hardware	Primary role
Cosmos 3 Edge	4B	Jetson Thor T2000 / T3000	Real-time on-device inference, robot policy
Cosmos 3 Nano	16B	RTX PRO 6000-class workstation GPUs	Fast video and action reasoning
Cosmos 3 Super	64B	Hopper / Blackwell datacenter GPUs	Post-training, highest physics accuracy

NVIDIA has not published direct Edge-vs-Nano latency or accuracy comparisons. Base Cosmos 3 has been described by NVIDIA as ranking first among open models on physical AI benchmarks including Physics-IQ, PAI-Bench, R-Bench, RoboLab, and RoboArena — but those numbers are vendor-stated and were not independently reproduced at launch, and Edge-specific scores have not been released.

Availability: How to Access Cosmos 3 Edge

Access at the time of writing runs through NVIDIA's own channels:

• NVIDIA Cosmos family on Hugging Face — Nano and Super weights are already published on the NVIDIA Hugging Face organization.
• build.nvidia.com — the model garden where Cosmos 3 tiers can be tried in-browser.
• GitHub — the open Cosmos framework repository under the NVIDIA org, for post-training and adaptation.
• Jetson Thor — the target deployment hardware, with JetPack emulation available now and physical T2000 / T3000 modules shipping Q1 2027.

Cosmos 3 Edge is available through the open Cosmos framework as of the July 15, 2026 launch, but the specific Hugging Face publication date for Edge weights was not confirmed at the announcement. Nano and Super are downloadable today; Edge availability tracks the Jetson Thor timeline.

What We Don't Know Yet

Several material questions remain open:

• Edge-specific benchmarks. No latency, throughput, or accuracy numbers were published for the 4B Edge variant against Nano, Super, or third-party robot policies.
• Hugging Face weight publication. NVIDIA confirmed the Edge tier, its 4B size, and framework availability, but did not confirm a Hugging Face publication date for standalone Edge weights.
• Real deployment results. The Japanese coalition members (FANUC, Yaskawa, Kawasaki, Sony, SoftBank, Fujitsu, Hitachi, Kubota, NEC, AIRoA) have stated intent to join and, in Fujitsu's case, to lead a collaborative control platform — but no factory-floor pilot results have been reported.
• Comparative safety data. Independent safety or reliability data versus proprietary robotics stacks (e.g. from Waymo, Tesla, or humanoid vendors) has not been published.

Frequently Asked Questions

What is Cosmos 3 Edge?

Cosmos 3 Edge is a 4-billion-parameter open world model from NVIDIA, built on Nemotron, designed for real-time on-device vision reasoning and robot action generation on Jetson, RTX, and DGX hardware. It is the edge tier of the Cosmos 3 family and was launched July 15, 2026 in Tokyo, following the May 31, 2026 debut of Cosmos 3 Nano and Super at GTC Taipei.

How is Cosmos 3 Edge different from Cosmos 3 Nano and Super?

Cosmos 3 Edge is the smallest tier at 4B parameters, sized for Jetson-class edge GPUs and real-time on-device inference. Cosmos 3 Nano is 16B for workstation-class GPUs like the RTX PRO 6000, and Cosmos 3 Super is 64B for Hopper and Blackwell datacenter GPUs used in post-training runs.

Is Cosmos 3 Edge open source?

The Cosmos 3 family ships under the Linux Foundation OpenMDW-1.1 license, which permits commercial use and derivatives with a "Built on NVIDIA Cosmos" attribution. Cosmos 3 Nano and Super weights are on Hugging Face; Cosmos 3 Edge was announced as available through the open Cosmos framework, though the exact Hugging Face publication date for Edge weights was not confirmed at launch.

How much does Cosmos 3 Edge cost?

There is no per-token or per-call pricing. Cosmos 3 Edge is an open foundation model distributed under OpenMDW-1.1, so the practical cost is the NVIDIA hardware it targets — Jetson Thor T2000 and T3000 modules, RTX GPUs, or DGX systems.

What hardware does Cosmos 3 Edge run on?

Cosmos 3 Edge is deployable across NVIDIA RTX GPUs, DGX systems, and the Jetson lineup, including the newly announced Jetson Thor T2000 (400 FP4 TFLOPS, 16 GB) and T3000 (865 FP4 TFLOPS, 32 GB LPDDR5X) modules. Jetson Thor emulation is available now through JetPack, with the physical modules shipping in Q1 2027.

What can developers do with Cosmos 3 Edge?

Developers can adapt Cosmos 3 Edge in about a day for specific robots, vehicles, sensors, and environments to produce on-device vision reasoning and robot policy models. Its outputs include text, video, and native action signals such as joint angles, gripper positions, and trajectory points.

Cosmos 3 Edge vs a large language model — what is the difference?

A large language model consumes and emits tokens over language. Cosmos 3 Edge is a world model that also ingests RGB video, depth, point clouds, and motion sequences, and emits video frames, ambient sound, and direct robot control signals, which is what makes it usable as a robot policy backbone rather than a chatbot.

What to Watch Next

Three signals will tell you whether Cosmos 3 Edge lives up to its framing. First, publication of standalone Edge weights on Hugging Face — Nano and Super are already there; a discrete Edge model card is the missing artifact. Second, independent latency and accuracy benchmarks on Jetson Thor T2000 and T3000 once the modules ship in Q1 2027. Third, the first non-vendor pilot report from a coalition member such as FANUC, Yaskawa, or Fujitsu showing Cosmos 3 Edge running a real production task, not a demo.

Building similar on-device or generative multimodal workloads? On kie.ai you can try Seedance 2.5, Wan 2.7 Video, and Gemini Omni.