# NVIDIA Cosmos 3

> Source: https://aiwiki.ai/wiki/nvidia_cosmos_3
> Updated: 2026-08-20
> Fact-checked: 2026-08-20
> Categories: Generative AI, NVIDIA, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "NVIDIA Cosmos 3." aiwiki.ai, 20 Aug 2026. https://aiwiki.ai/wiki/nvidia_cosmos_3
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**NVIDIA Cosmos 3** is an open family of "world [foundation models](https://aiwiki.ai/wiki/foundation_models)" for [physical AI](https://aiwiki.ai/wiki/physical_ai) that [NVIDIA](https://aiwiki.ai/wiki/nvidia) launched on June 1, 2026 at GTC Taipei, held alongside COMPUTEX 2026. NVIDIA describes it as the first fully open "omnimodel," meaning a single model that can both reason about the physical world and generate it across text, images, video, ambient sound and action, which the company says reduces physical-AI training and evaluation cycles "from months to days."[^1][^2][^3] Cosmos 3 is built on a Mixture-of-Transformers (MoT) architecture that pairs a reasoning transformer with an expert generation transformer, so the model can understand object interactions, motion and spatial-temporal relationships before it generates video and action trajectories.[^1][^2][^3] It is the third major generation of the [NVIDIA Cosmos](https://aiwiki.ai/wiki/nvidia_cosmos) platform, which NVIDIA positions as the data and simulation layer for [robots](https://aiwiki.ai/wiki/robot_learning), [autonomous vehicles](https://aiwiki.ai/wiki/autonomous_vehicle) and other machines that perceive and act in the real world.

The launch consolidated what had previously been several separate Cosmos models into one. NVIDIA also used the announcement to introduce the **NVIDIA Cosmos Coalition**, a group of AI labs and robotics companies committed to advancing open [world models](https://aiwiki.ai/wiki/world_model).[^1][^4]

## Key facts

| Attribute | Detail |
| --- | --- |
| Developer | [NVIDIA](https://aiwiki.ai/wiki/nvidia) (Cosmos research lab) |
| Announced | June 1, 2026, at GTC Taipei / COMPUTEX 2026 ([Jensen Huang](https://aiwiki.ai/wiki/jensen_huang) keynote); Edge added at SIGGRAPH 2026 (July 20, 2026) |
| Type | Open world foundation model / "omnimodel" for physical AI |
| Architecture | Mixture-of-Transformers (autoregressive reasoner tower + diffusion generator tower) |
| Variants | Cosmos 3 Super (64B), Cosmos 3 Nano (16B), Cosmos 3 Edge (4B) |
| Family modalities | Text, image, video, ambient sound (audio), action |
| License | OpenMDW-1.1 (Open Model, Data and Weights License), Linux Foundation |
| Availability | build.nvidia.com, Hugging Face, GitHub; deployable as [NVIDIA NIM](https://aiwiki.ai/wiki/nvidia_nim) microservices |
| Released checkpoints | nvidia/Cosmos3-Nano, nvidia/Cosmos3-Super (May 31, 2026); nvidia/Cosmos3-Edge (July 20, 2026); task checkpoints including Cosmos3-Nano-Policy-DROID and Cosmos3-Edge-Policy-DROID |
| Tested precision | BF16 (FP4, FP8 and FP16 not officially supported) |
| Supported GPUs | NVIDIA Ampere, [Hopper](https://aiwiki.ai/wiki/nvidia_hopper) and [Blackwell](https://aiwiki.ai/wiki/nvidia_blackwell); Edge also runs on Jetson Thor and the Jetson T2000 / T3000 edge modules |

## What is NVIDIA Cosmos 3?

Most large language models learn from text written from a human point of view. Robots and self-driving cars need something different: data captured from their own perspective, including how objects move, how surfaces behave and what happens when a gripper closes on something. In his keynote, [Jensen Huang](https://aiwiki.ai/wiki/jensen_huang) called this one of the hardest data problems in computing, because the real world rarely hands you labeled examples of every edge case a machine might encounter.[^2] Cosmos 3 is NVIDIA's attempt to attack that problem with a single model that can understand a scene and then synthesize new, physically grounded versions of it.

NVIDIA frames Cosmos 3 as an "omnimodel" because it folds together capabilities that used to require a small stack of specialized models. Earlier in the Cosmos line, developers worked with separate components: Cosmos Predict for world generation, Cosmos Transfer for controlled generation, Cosmos Reason for scene understanding, and Cosmos Policy for action and policy generation. Cosmos 3 does all of that in one model that can reason and generate across modalities in a unified forward pass, which removes a lot of the orchestration glue that linking four models together normally demands.[^3][^5] For the history of those earlier components and the broader platform, see the main [NVIDIA Cosmos](https://aiwiki.ai/wiki/nvidia_cosmos) article.

The headline pitch is "think before it acts." The model first interprets what is happening in a scene, then uses that understanding to produce outputs, whether that output is a predicted video of what comes next or a set of numerical actions for a robot to execute.[^2] Ming-Yu Liu, NVIDIA's vice president of research who leads the Cosmos Lab, describes the design as one that "first harnesses a reasoning block to interpret what is happening in a scene, then harnesses a generation block to use that context to create physically grounded outputs."[^2] Jensen Huang summed up the ambition this way: "The big bang of physical AI is just around the corner thanks to breakthroughs in multimodal reasoning language, vision and world models. The Cosmos 3 family of open, frontier omnimodels gives developers a generational leap in ability to build robots, autonomous vehicles and vision AI that perceive, reason, plan and act in the physical world."[^1]

## What is a world foundation model?

A world foundation model (WFM) is a model pretrained to understand and generate the dynamics of the physical world, so that other systems can be built on top of it the way applications are built on a [large language model](https://aiwiki.ai/wiki/large_language_model). NVIDIA positions Cosmos as a WFM platform whose job is to supply the scarce ingredient in robotics and autonomy: realistic, physics-aware data. Rather than collecting and labeling every real-world scenario, developers can ask a WFM to generate diverse, physically plausible scenes, including rare edge cases, and use that synthetic output to train and evaluate their machines.[^4] What makes Cosmos 3 unusual among WFMs is that it unifies world generation, physical reasoning and action generation inside one open model, where earlier approaches kept those as separate models.[^5]

## Mixture-of-Transformers architecture

The technical core of Cosmos 3 is a Mixture-of-Transformers (MoT) backbone built from two complementary "towers" that live inside one model.[^3][^6]

The first is the **Reasoner**, an autoregressive [vision-language model](https://aiwiki.ai/wiki/vision_language_model). It reads multimodal observations such as images, video and text (with a [context window](https://aiwiki.ai/wiki/context_window) of up to 256K tokens) and produces text: [captions](https://aiwiki.ai/wiki/captions_ai), plans, spatial and temporal reasoning, and judgments about physical plausibility. It works the way a [language model](https://aiwiki.ai/wiki/language_model) does, predicting the next token in a sequence, which lets it describe motion, object interactions and other physical context.[^3][^6][^8]

The second is the **Generator**, a [diffusion](https://aiwiki.ai/wiki/diffusion_model) transformer. Conditioned on the Reasoner's understanding, it produces the continuous, non-text outputs: images, video, audio and action sequences, generated through iterative [denoising](https://aiwiki.ai/wiki/denoising) rather than token-by-token decoding. In practice that means the Generator can handle text-to-image, text-to-video, image-to-video, video-to-video, audio synthesis and action generation, including forward dynamics, inverse dynamics and policy outputs for robots.[^3][^6]

What makes this a single model rather than two stitched together is how the towers share a sequence. NVIDIA's technical report describes splitting the input into two subsequences: an autoregressive (AR) subsequence that handles reasoning via next-token prediction, and a diffusion (DM) subsequence that handles generation via iterative denoising. The AR and DM tokens use separate parameter sets within each transformer layer, but they interact through joint attention, and information flows one way, from the Reasoner to the Generator. So the two towers keep their own weights yet can attend to each other inside the same forward pass.[^7][^10] A 3D multimodal [rotary position embedding](https://aiwiki.ai/wiki/rotary_position_embedding) (mRoPE) aligns video, audio and action tokens on a single temporal axis so the model can keep them in sync.[^10] Every modality first passes through a dedicated encoder, a vision transformer (ViT) for visual understanding, a variational autoencoder (VAE) for visual and audio generation, and domain-aware vectors for actions, before being projected into a shared representation space.[^8]

This design is a deliberate bet against forcing every output through the same decoder. Reasoning and language are well suited to autoregressive prediction; high-fidelity pixels, audio and smooth trajectories are better served by diffusion. By keeping both mechanisms in one framework, Cosmos 3 can reason and generate heterogeneous modalities without pretending they should all be produced the same way.

### What are the inputs and outputs?

The two towers expose different interfaces. The published model card for Cosmos 3 Nano lists the following.[^8]

| Tower | Inputs | Outputs |
| --- | --- | --- |
| Reasoner (understanding) | Text; text plus image; text plus video (up to 256K-token context) | Text (reasoning, planning, spatial and temporal reasoning) |
| Generator (generation) | Text; image; video (with or without audio); action trajectory | Image (JPG), video (MP4), audio (stereo AAC, 48 kHz), action (JSON), text |

Action data is represented numerically, for example as joint angles, gripper positions and trajectory points, which is the format a robot controller can actually consume.[^2] On the action side, Cosmos 3 supports multiple embodiments, including camera, vehicle, egocentric (first-person), single-arm, dual-arm and humanoid robots.[^10]

## What are the Cosmos 3 model variants?

NVIDIA released Cosmos 3 in a tiered family so the same architecture can run from a workstation up to a data center, with a third tier, Cosmos 3 Edge, extending the family down to real-time, on-device inference.[^1][^9][^11]

| Variant | Total parameters | Tower split | Positioning | Status |
| --- | --- | --- | --- | --- |
| Cosmos 3 Super | 64B | ~32B reasoner + ~32B generator | Highest physics accuracy; for post-training robotics and AV models | Available |
| Cosmos 3 Nano | 16B | ~8B reasoner + ~8B generator | High-quality video and action reasoning in fractions of a second | Available |
| Cosmos 3 Edge | 4B | ~2B reasoner + ~2B generator | Real-time, on-device inference for robot policy and vision analytics | Available |

Cosmos 3 Super targets data-center [Hopper](https://aiwiki.ai/wiki/nvidia_hopper) and [Blackwell](https://aiwiki.ai/wiki/nvidia_blackwell) GPUs, while Nano is sized for workstation-class hardware such as the RTX PRO 6000 Blackwell card.[^5][^9] NVIDIA's testing was done at BF16 precision on GB200 and H100 systems; the model card notes that FP4, FP8 and FP16 are not officially supported at launch.[^8]

Beyond the two base models, the GitHub release also includes task-specialized checkpoints in the same family, such as text-to-image and image-to-video variants of Super and a robot policy model (Cosmos3-Nano-Policy-DROID) trained for manipulation, described in the world action model section below.[^6]

## Cosmos 3 Edge

NVIDIA completed the tiered family on July 20, 2026 with **Cosmos 3 Edge**, a 4-billion-parameter open omnimodel built for memory-efficient, high-throughput inference on edge devices.[^11][^24] Where Super and Nano target data-center and workstation hardware, Edge is sized to bring frontier world models onto the machine itself, so a robot policy or a video-analytics agent can run without a round trip to a data center. NVIDIA frames the need plainly: machines operate at the edge in factories, warehouses and hospitals, and to understand and act in those environments they need models that deliver data-center-level performance on memory-constrained systems.[^11]

Edge keeps the two-tower design of the rest of the family: an autoregressive transformer for understanding and reasoning, and a diffusion transformer for prediction and generation, joined by shared multimodal attention.[^7][^11] NVIDIA's technical report describes two dense towers of about 2 billion parameters each, with independent reasoner and generator parameter sets inside each Mixture-of-Transformers layer. The 2B reasoner uses a custom Nemotron backbone; the report says the Edge transformer was trained from scratch and largely follows the Qwen3-1.7B design, but it does not identify a specific pretrained Nemotron checkpoint.[^7][^12] NVIDIA's SIGGRAPH post says Edge handles the same five modalities as Super and Nano: text, image, video, ambient sound and action.[^12] The released artifacts are narrower. The Edge-specific model-card interfaces and repository matrix list text, image, video and action, while the checked-in configuration disables sound generation. As of the public release, audio is therefore a family-level launch claim rather than an exposed capability of the released Edge checkpoint.[^6][^13][^25] The weights carry the same **OpenMDW-1.1** license as the larger models, and only BF16 precision is tested; FP4, FP8 and FP16 are not officially supported.[^13]

### Hardware targets and on-device performance

Edge is built to run across NVIDIA's local-compute lineup rather than in the data center: [NVIDIA Jetson](https://aiwiki.ai/wiki/nvidia_jetson) modules (including the Jetson AGX Thor and the newly announced mid-range Jetson T3000 and T2000), RTX PRO GPUs, GeForce RTX GPUs, and DGX systems including [DGX Spark](https://aiwiki.ai/wiki/nvidia_dgx_spark) and DGX Station.[^11][^13] The T3000 and T2000 themselves were unveiled in NVIDIA's July 15, 2026 Jetson Thor announcement as Blackwell-based modules offering 865 and 400 FP4 teraflops of compute at 70 W and 40 W respectively, aimed at mainstream robotics and [edge AI](https://aiwiki.ai/wiki/edge_ai), with availability slated for the first quarter of 2027.[^15] That announcement already pointed to Cosmos 3 Edge as "a 4-billion-parameter model helping embodied systems see the world, reason over it in real time, and predict and generate actions through on-device inference," which is why some coverage dates the model to July 15; the open weights and model card were published on July 20 alongside the SIGGRAPH keynote.[^13][^15]

NVIDIA's August tutorial and current policy card give an exact PyTorch test for the DROID checkpoint. Each loopback request used one persistent connection and no concurrency, a three-camera canvas 540 pixels high by 640 wide, a 544 by 736 processing bucket, four UniPC denoising steps, guidance 3.0 and a prediction shaped `[32, 8]` at `conditioning_fps=15`.[^14][^41] The 32 rows are future action timesteps and the eight values are seven absolute joint positions plus one gripper value. At 15 Hz, a chunk represents 2.133 seconds of commanded motion, so the card defines its real-time factor as 2.133 seconds divided by measured end-to-end latency.[^14]

| Jetson configuration | Median E2E | Observed range | Timed requests | RTF at 15 Hz | Meets the card's 15 Hz budget |
| --- | ---: | ---: | ---: | ---: | --- |
| AGX Thor T5000, 128 GB, MAXN, 1575 MHz | 1.528 s | 1.503 to 1.673 s | 16 | 1.40 | Yes |
| AGX Thor T4000, 64 GB class, MAXN, 1530 MHz | 2.208 s | 2.148 to 2.425 s | 9 | 0.97 | No |
| Thor T3000, 32 GB, 1100 MHz, emulated | 2.632 s | 2.595 to 2.687 s | 4 | 0.81 | No |
| Thor T2000, 16 GB, 765 MHz, `THOR_NANO` | 5.195 s | 5.176 to 5.224 s | 9 | 0.41 | No |

Only the T5000 met this particular 15 Hz budget. The T4000's fastest observed request was still 0.7% over it. NVIDIA emulated the T3000 with memory, SM, GPU-clock and CPU caps but left DRAM bandwidth uncapped, so production hardware may be slower. The T2000 run used 2.6 GB of swap with other Cosmos containers stopped, although the card says no swap-out occurred during the timed calls.[^14] These are NVIDIA measurements of policy-server latency, not independent end-to-end robot tests. They do not include a deployed robot's full sensor, safety-controller, actuator and physical-response path.

The same card contains a separate vLLM-Omni test with 320 by 192 observations, 30 denoising steps, guidance 1.0 and a 5 Hz budget. It used one discarded warmup and two timed runs per row. The T5000, T4000 and emulated T3000 met that looser 6.4-second chunk budget, while the T2000 did not.[^14] The vLLM and PyTorch numbers should not be compared directly because their inputs, denoising settings, runtime stacks and timing protocols differ.

### Benchmarks and the DROID policy

NVIDIA's benchmark claims for Edge are its own. The company says that among similar-size (4B-parameter) models, Cosmos 3 Edge "ranks #1 on VANTAGE-Bench for vision analytics and state-of-the-art for robot policy learning."[^11][^12] The current model card now includes the vendor's VANTAGE-derived reasoning table: it reports a 50.3 average for Edge on the smart-infrastructure group, compared with 47.7 for Cosmos-Reason2-2B and 42.5 for Qwen3-2B-VL, although results across the individual tasks are mixed. Its generation table reports 81.2 on PAIBench, 51.9 on RBench and 33.2 on PhysicsIQ, while the policy card reports RoboLab success rates that vary with instruction specificity.[^13][^14] All of these evaluations were run or reported by NVIDIA. Without independent testing, the leaderboard language remains a vendor claim rather than evidence of uniform superiority.

Alongside the base model, NVIDIA released a task-specialized checkpoint, **Cosmos 3 Edge Policy (DROID)** (nvidia/Cosmos3-Edge-Policy-DROID), a [robot-manipulation](https://aiwiki.ai/wiki/robot_manipulation) policy post-trained on the DROID dataset. It turns a language instruction, three camera views and measured robot state into an action trajectory, and its documented output is DROID's 8D absolute joint-position interface rather than a universal controller for arbitrary embodiments.[^14][^41] NVIDIA illustrates the policy with a Franka Panda arm and Robotiq gripper, while its August tutorial says another embodiment needs its own action space, camera layout, normalization and experiment configuration.[^26][^41] NVIDIA presented Cosmos 3 Edge at its SIGGRAPH 2026 keynote in Los Angeles on July 20, 2026, delivered by Ming-Yu Liu, the vice president who leads NVIDIA's Cosmos research lab.[^12][^17]

The DROID source counts depend on the artifact. The full text of the revised DROID paper says the original release contains 76,000 successful episodes plus roughly 16,000 unsuccessful trajectories that were released but excluded from the headline count.[^44] Its main text and Figure 1 use 86 task verbs, while the arXiv metadata abstract says 84 tasks. NVIDIA's current LeRobot conversion instead reports 57,639 episodes in its `success/` split and 14,268 in `failure/`, or 71,907 total episodes at 15 Hz.[^43] The post-training recipe reads the converted success split and then applies a keep-ranges filter that removes idle and non-task windows.[^42] The tutorial's phrase "76K successful" therefore matches the original DROID release, but it is not the episode count of the converted training split.

RoboLab-120 is a closed-loop simulation benchmark with 120 language-conditioned manipulation tasks. Its paper describes visual, procedural and relational competency axes and reports that task-generalist policies remain sensitive to instruction wording, scene complexity and task horizon.[^45] NVIDIA's current model-card figure gives the following overall success rates:[^14]

| Policy | Vague instructions | Default instructions | Specific instructions |
| --- | ---: | ---: | ---: |
| Cosmos3-Edge-Policy-DROID | 15.4% | 22.9% | 28.8% |
| Cosmos3-Nano-Policy-DROID | 20.6% | 36.8% | 39.7% |

The tutorial's 22.9% headline is the Edge policy's overall result with the benchmark's default instructions. It is not a real-robot success rate, and the same figure reports 11.8% on complex tasks under that instruction condition.[^14] RoboLab is maintained by NVIDIA and the policy results have not been independently reproduced.

### Reproducible DROID post-training workflow

NVIDIA publishes a recipe-level reproduction workflow. A developer downloads Cosmos3-DROID, converts the Cosmos 3 Edge base checkpoint to distributed-checkpoint format, enables the published keep-ranges filter, substitutes `EDGE_MODEL_CONFIG` for the repository's Nano configuration, points the recipe at the Edge checkpoint and launches supervised post-training.[^41][^42] The conversion step repacks weights on CPU; it is not model training. The action encoder, decoding MLP and action embedding parameters are initialized separately from the base checkpoint.[^42]

| Training setting | Published value |
| --- | --- |
| Action and state | 8D absolute joint position, with proprioception enabled |
| Observation | Three-camera, 540-high by 640-wide canvas |
| Prediction | 32 future actions at 15 Hz |
| Image preparation | Random crop, rescale and color jitter |
| Joint normalization | None; raw values |
| Learning rate | `2e-4` |
| Global batch | 8,192 samples across 256 ranks in the reference layout |
| HSDP layout | 32 replicas by 8-way shard |
| Action-head learning-rate multiplier | 5x |
| Vision and action loss weights | 10 each |
| Public schedule | 10,000 iterations in a 100,000-step cosine cycle |
| Checkpoint interval | 1,000 iterations |

NVIDIA describes this as foundation-model post-training, not a single-GPU fine-tune. The reference HSDP layout corresponds to 64 nodes with four GB200 GPUs each. The public TOML and tutorial table specify 10,000 iterations, but the same tutorial separately calls 60,000 iterations on that hardware for about 68 hours, or roughly 17,400 GB200 GPU-hours, the "validated run."[^41][^42] The sources do not reconcile the 10,000- and 60,000-iteration figures, so the latter is a vendor compute disclosure rather than the unambiguous public reproduction recipe.

The model weights, converted dataset and Cosmos framework are released under OpenMDW-1.1, while RoboLab's framework code uses Apache 2.0.[^14][^27][^43][^47] RoboLab's bundled assets carry separate folder-specific terms, including Creative Commons licenses and some noncommercial restrictions.[^47] OpenMDW-1.1 permits use, modification and redistribution of the model materials, but redistribution must retain the license and applicable notices, and the agreement contains a patent-and-copyright-litigation termination clause.[^27] The tutorial expects CUDA 13.0, NVIDIA's 26.06 container stack, access to the base checkpoint and dataset, and a Hugging Face token in its command examples.[^41] As of August 20, the base, policy and converted-dataset repositories were publicly readable without an access gate.[^13][^14][^43] Public availability is not a safety certification, and the checkpoint still needs use-case-specific validation.

### Receding-horizon serving and limitations

NVIDIA says the BF16 deployment occupies roughly 9 GB of weights. Its Thor example runs an OpenPI-compatible WebSocket policy server locally, sets `TORCHDYNAMO_DISABLE=1` for eager execution because the tutorial says stock Triton wheels lack `sm_110a` kernels, then sends a language instruction, three images, seven joint positions and one gripper position for each query.[^41][^42] The zero-filled state in the tutorial is explicitly a smoke test; a robot loop must supply measured joint and gripper state.

The tutorial's schematic receding-horizon loop predicts 32 actions, executes the first 16 at 15 Hz, captures a new observation and queries again. That prefix spans about 1.067 seconds. Predicting a longer chunk and executing a shorter prefix before replanning follows the receding-horizon approach used in diffusion-based robot policies to balance temporal consistency against responsiveness.[^41][^46]

That schematic should not be confused with a measured asynchronous controller. The tutorial pseudocode waits for inference before executing its prefix, the card's timing test sends one request at a time, and the current RoboLab client consumes all 32 cached actions before making its next synchronous request.[^14][^41][^47] The fact that the T5000's 1.528-second median is shorter than the full 2.133-second chunk is therefore a throughput-budget result, not proof that the published client overlaps inference with execution or that a physical arm moves without pauses.

The model card also warns that Cosmos 3 can show temporal inconsistency, imprecise physical interactions, incorrect reasoning and action-state drift, especially on long-horizon or out-of-distribution inputs. It says the outputs are not physically accurate simulation, ground-truth reasoning or safety-certified decisions.[^14] No independent full training reproduction or real-robot evaluation of the released Edge DROID checkpoint was found as of August 20, 2026.

### Edge deployment and evaluation partners

Separately from the Cosmos Coalition founders listed further below, NVIDIA's SIGGRAPH announcement named a set of early partners putting Cosmos 3 Edge through evaluation, split between robotics workflows and edge vision analytics.[^12] On the robotics side, [Agile Robots](https://aiwiki.ai/wiki/agile_robots), [Doosan Robotics](https://aiwiki.ai/wiki/doosan_robotics), [Siemens](https://aiwiki.ai/wiki/siemens) and [Skild AI](https://aiwiki.ai/wiki/skild_ai) are named among the partners evaluating Cosmos 3 Edge for robotics workflows.[^12] On the vision side, Centific, Vaidio and YUAN are evaluating it to accelerate vision agents running at the edge.[^12] These are launch-time evaluation or self-reported integration statements, not independent evidence that Edge has been deployed in production. The launch groups the companies with a single line each, but their own materials describe fairly different uses, so it is worth taking them one at a time rather than lumping them together.

Vaidio describes Cosmos 3 Edge as a second-stage reasoning layer in its Vaidio Vision AI platform rather than as a first-pass detector. In Vaidio's account, object detection runs first, then only the relevant image or crop is sent to Edge for an extra verification step that discards the event or promotes it to an alert. The company says it combines this pipeline with NVIDIA Metropolis libraries on Jetson Thor for remote industrial sites, mobile systems and robot-mounted cameras.[^22] This is Vaidio's own integration description, not an independently audited deployment. Its write-up is titled around "Cosmos Reason 3 Edge" even though the body calls the model "Cosmos 3 Edge," a framing that reflects its use of the reasoner rather than the generation tower.[^22]

Centific says it completed a standalone evaluation of Cosmos 3 Edge and is bringing the compact 4B model, including its 2B reasoner, into production pipelines for video understanding. In Centific's own single-H100 BF16 test, a frame-level answer took about 2.3 seconds, time to first token was about 40 ms, throughput was about 57 tokens per second, GPU memory stayed under 6 GB, and a 16-frame clip completed in under 6 seconds.[^23] Centific cautions that embedded latency varies, so these partner-reported H100 figures are not Jetson measurements or independent production evidence. YUAN is likewise evaluating Edge for edge vision agents and sits among NVIDIA's Jetson Thor ecosystem partners, though NVIDIA's blog is the main public source for its involvement and does not detail a specific deployment.[^12]

## Cosmos-Dreams: closed-loop AV simulation at SIGGRAPH

At the same SIGGRAPH 2026 keynote where it launched Cosmos 3 Edge, NVIDIA also showed **Cosmos-Dreams**, described as a collection of closed-loop simulators built on NVIDIA Cosmos world models.[^12] The headline demonstration was a simulator built for autonomous vehicles that generates an entire driving world from a single frame and lets a driver move through a scene that never existed in reality, rendered frame by frame in real time on a single [NVIDIA RTX PRO 6000](https://aiwiki.ai/wiki/nvidia_rtx_pro_6000) GPU.[^12]

Ming-Yu Liu, the vice president who leads NVIDIA's Cosmos Lab, presented the segment, and an NVIDIA colleague drove the live demo using a PlayStation 5 controller, with the simulator generating photorealistic, frame-by-frame output in response to the vehicle's actions.[^12][^21] NVIDIA presented the efficiency as a steep descent: the presenter said earlier versions of the simulator had needed 64 GB300 GPUs, then 16 of NVIDIA's newer Vera Rubin GPUs, before the current build was optimized down to that single workstation card.[^21] One naming detail is worth pinning down, since the demo and the blog disagree: in the live keynote the presenter referred to the card as an "RTX 6000 Ada Generation" workstation GPU, while NVIDIA's SIGGRAPH blog and the underlying omni-dreams model card both point instead to the Blackwell-generation RTX PRO 6000 (the card the model card lists as "RTX 6000 Pro"), so the blog wording is used above.[^12][^19][^21]

NVIDIA's pitch for Cosmos-Dreams is the same closed-loop story as the rest of the platform: developers can use it to verify the accuracy of their models before deploying them to a real fleet, or to train models on AI-generated scenarios that are hard to capture in the real world, which the company says accelerates development while saving on operational costs.[^12] That fits the same autonomous-vehicle stack as NVIDIA's open driving models, such as [NVIDIA Alpamayo 2 Super](https://aiwiki.ai/wiki/nvidia_alpamayo_2_super), which are meant to be exercised against generated worlds before they meet real traffic.

Cosmos-Dreams is the renamed continuation of NVIDIA's earlier **OmniDreams** research. NVIDIA's Toronto AI Lab publishes the project on GitHub with the description "NVIDIA Cosmos-Dreams (fka NVIDIA OmniDreams) is a world model that generates photorealistic video for autonomous-driving simulation in real time," where "fka" marks the rename.[^18][^20] The underlying research model was released weeks before SIGGRAPH, on June 2, 2026, as the open "omni-dreams" models on GitHub and Hugging Face: a roughly 2-billion-parameter single-view world model, developed on top of NVIDIA's Cosmos-Predict 2.5, that takes a single input frame plus a text prompt, HD-map images and trajectory poses and produces camera video autoregressively for closed-loop driving simulation.[^18][^19] The code is offered under the Apache 2.0 license and the weights under NVIDIA's Open Model License Agreement.[^18][^19]

One point of nomenclature is worth keeping straight. Cosmos-Dreams, the autonomous-vehicle simulator, is not the same thing as GR00T-Dreams, an earlier NVIDIA robotics data blueprint that also uses Cosmos world models but generates synthetic "dream" training data for humanoid robots. GR00T-Dreams is covered on the main NVIDIA Cosmos page and predates the Cosmos-Dreams AV work described here.

## How is Cosmos 3 trained, and on what data?

Cosmos 3 is trained on large-scale multimodal data rather than text alone. According to the Cosmos 3 Nano model card, the model was trained on roughly 1.3 billion data points drawn from 393 dataset entries collected between 2024 and 2026. The breakdown spans the modalities the model has to master: about 22 million reasoning (text) samples, 767 million images, 348 million videos, 139 million audio samples and 8 million action samples.[^8]

The two towers learn in complementary ways. The Reasoner is trained on paired vision-language data such as image-text and video-text pairs to support question answering, spatial grounding, temporal reasoning and action understanding. The Generator is trained on large multimodal corpora of images, video, audio and action using reconstruction-based objectives, so it learns to synthesize the world rather than to label it.[^7] To address the shortage of robot-perspective data, NVIDIA says Cosmos 3 also learns from teleoperation recordings, simulation, and third-person video that is re-projected into a first-person view.[^4]

A widely repeated figure of "20 trillion training tokens" circulated in some press coverage of the launch, but that number does not appear in NVIDIA's official model cards, technical blog or technical report, so it is omitted here. NVIDIA's own documentation describes the data in terms of samples per modality, as above, rather than a single token count.

## What can Cosmos 3 do?

NVIDIA presents Cosmos 3 as a leaderboard-topping open model across the three things it is built to do: physical-AI reasoning, world simulation and action generation. Performance claims below are NVIDIA's own.

The company says Cosmos 3 ranks first among open models on a spread of physical-AI and generation leaderboards, including Artificial Analysis (for text-to-image and image-to-video), Physics-IQ, PAI-Bench, R-Bench, RoboLab, RoboArena, VANTAGE-Bench and TAR.[^1] NVIDIA's developer materials add that Cosmos 3 leads VANTAGE-Bench at both the 32B tier (Super) and the 8B tier (Nano), framing the comparison by the size of the individual towers rather than the combined model; with Edge, NVIDIA extends that VANTAGE-Bench claim down to the 4B class.[^5][^11] NVIDIA has not published a full table of numerical scores in its launch blogs; the detailed results live in the accompanying technical report.[^7]

NVIDIA restated the rankings in an August 6, 2026 corporate blog post by Ming-Yu Liu: Cosmos 3 ranks No. 1 on Artificial Analysis for [open-weights](https://aiwiki.ai/wiki/open_weights) text-to-image and image-to-video generation, No. 1 on PAI-Bench for world generation and in the image-to-video category of Physics-IQ, and No. 1 on RoboLab for robot policy, while Cosmos 3 Super is the highest-ranked open model on VANTAGE-Bench for vision understanding.[^35] The live Artificial Analysis arenas back only part of that. As of August 8, 2026, Cosmos3-Super-Text2Image (agentic) did lead the open-weights field in the Text to Image Arena with an Elo of 1236, but the open-weights lead in the Image to Video Arena had passed to [MiniMax H3](https://aiwiki.ai/wiki/minimax_h3), at an Elo of 1351 in the without-audio ranking, with Cosmos3-Super-Image2Video-4Step second among open weights at 1265.[^36][^37]

Practically, the model is aimed at a few concrete jobs. It can generate [synthetic training data](https://aiwiki.ai/wiki/synthetic_data) and rare edge cases for robots and self-driving systems, predict how a scene will unfold, reason about what a camera is seeing for video-analytics agents in factories and cities, and output the actual control actions for robot manipulation such as pick-and-place.[^2][^9] Because the Reasoner and Generator share one model, a developer can ask Cosmos 3 to look at a scene, decide what should happen, and then produce both the predicted video and the action trajectory to get there. NVIDIA's pitch is that this compresses the physical-AI development loop "from months to days," by replacing demonstrations that would otherwise have to be captured by hand.[^1][^2]

## Is Cosmos 3 open source, and how do you get it?

Cosmos 3 Super and Cosmos 3 Nano launched on May 31, 2026, and Cosmos 3 Edge followed on July 20, 2026, so all three tiers are now available.[^1][^11] The models can be tried as hosted endpoints on build.nvidia.com, downloaded as open weights from Hugging Face (the nvidia/Cosmos3-Nano and nvidia/Cosmos3-Super collections, both dated May 31, 2026, and nvidia/Cosmos3-Edge, dated July 20, 2026), and built on through the NVIDIA Cosmos repository on GitHub, which also ships training and post-training scripts.[^1][^6][^8][^13] For production, the models can be deployed as [NVIDIA NIM](https://aiwiki.ai/wiki/nvidia_nim) microservices, and NVIDIA names cloud and inference partners including Baseten, CoreWeave, Microsoft Azure, Nebius, Deep Infra and Classmethod for hosting them.[^1]

The weights are released under the **OpenMDW-1.1** license, the **Open Model, Data and Weights License**, stewarded by the Linux Foundation. Its terms permit commercial and noncommercial use, modification and distribution of the model materials subject to retaining license and copyright notices; they include patent and copyright litigation termination provisions and do not restrict model outputs.[^13][^27][^28] NVIDIA also lists a contact address (cosmos-license@nvidia.com) for custom arrangements.[^6]

The licensing choice lines up with a wider policy position. NVIDIA signed "Open Weights and American AI Leadership," a July 24, 2026 open letter arguing that US AI leadership will be judged not by any single frontier model but by whether a strong, open ecosystem diffuses into every sector; the version of the letter published on NVIDIA's site listed 235 company and organization signatories as of August 8, 2026.[^38] Ming-Yu Liu's August 6, 2026 blog post applied that argument to physical AI, where he wrote that every deployment is a specialization problem, so closing the gap between a general model and a particular robot requires access to the weights, a license that permits adaptation and post-training tools: "Specialization is where openness becomes a practical technical requirement."[^35]

Developers typically adapt Cosmos 3 to their own robots or vehicles through supervised fine-tuning on custom datasets and through action post-training that covers forward dynamics, inverse dynamics and policy generation. NVIDIA's tooling supports quantization (BF16, FP8 and NVFP4) for NIM deployment and integration with the vLLM inference engine for higher throughput, although the base checkpoints themselves were validated at BF16.[^5]

## What is the NVIDIA Cosmos Coalition?

Alongside the model, NVIDIA announced the **NVIDIA Cosmos Coalition**, described as a group of leading AI labs and robotics companies working together to advance the next generation of open world models. The founding members named at launch are Agile Robots, Black Forest Labs, Generalist, LTX, Runway and Skild AI.[^1][^4] The grouping is notable for mixing robotics specialists (Agile Robots, Generalist, Skild AI) with generative-media labs (Black Forest Labs, LTX, Runway), which fits Cosmos 3's dual nature as both a robot-action model and a high-fidelity video generator.

NVIDIA expanded the coalition to Japan on July 15, 2026, in the same announcement that introduced Cosmos 3 Edge and the new Jetson Thor modules. Japanese physical-AI companies including AIRoA, classmethod, Enactic, [FANUC](https://aiwiki.ai/wiki/fanuc), Fujitsu, GROOVE X, Hitachi, Honda R&D, Kawasaki Heavy Industries, Kubota, Mitsui & Co., Mitsubishi Corp., Mujin, NEC, Preferred Networks, SoftBank Corp., [Sony Group Corporation](https://aiwiki.ai/wiki/sony), Telexistence, TIER IV, TRON K.K., Turing and [Yaskawa Electric](https://aiwiki.ai/wiki/yaskawa) intend to join, NVIDIA said, with the resulting world models meant to shorten development cycles across factories, logistics networks, farms, construction sites, hospitals, roads and homes.[^39]

An August 6, 2026 blog post rounding up the open-model push listed adopters across the wider Cosmos family, beyond the Edge evaluation partners named above: Doosan Robotics, [LG Electronics](https://aiwiki.ai/wiki/lg_electronics), [Samsung Electronics](https://aiwiki.ai/wiki/samsung) and Skild AI building on Cosmos in robotics; Li Auto, [Xiaomi](https://aiwiki.ai/wiki/xiaomi) and Afari, the English brand of the Chinese smart-driving company Chongqing Qianli Technology, in autonomous vehicles; and Centific, Fogsphere, Linker Vision, Milestone Systems and Yuan in vision AI agents for industrial AI and smart-spaces applications.[^35][^40] These are NVIDIA's adoption statements rather than announcements from the companies themselves.

## Where does Cosmos 3 fit in NVIDIA's physical-AI stack?

Cosmos 3 is the model layer in a larger physical-AI strategy NVIDIA has been assembling for several years. The company connects it to its simulation and robotics software, [NVIDIA Omniverse](https://aiwiki.ai/wiki/nvidia_omniverse) and the Isaac platform, where synthetic worlds are built and robot policies are tested before deployment. On the robotics side it lines up with [Isaac GR00T](https://aiwiki.ai/wiki/isaac_gr00t), NVIDIA's foundation-model effort for humanoid robots, and with simulation tools like [Isaac Sim](https://aiwiki.ai/wiki/nvidia_isaac_sim) and [Isaac Lab](https://aiwiki.ai/wiki/isaac_lab). For autonomous vehicles it connects to the NVIDIA DRIVE Hyperion platform, and for on-robot inference it pairs with the [Jetson Thor](https://aiwiki.ai/wiki/jetson_thor) edge computer.[^2]

That positioning maps onto NVIDIA's "three computers" view of robotics: one computer (in the data center, on [Blackwell](https://aiwiki.ai/wiki/nvidia_blackwell) and the upcoming [Vera Rubin](https://aiwiki.ai/wiki/nvidia_vera_rubin) generation) to train the models, a second running Omniverse and Cosmos to generate data and simulate, and a third (Jetson Thor) inside the robot to run the policy in the real world. Cosmos 3 is designed to feed the middle of that loop, turning scarce real-world data into the abundant, physically grounded data that the other two computers need, while Cosmos 3 Edge now supplies a model small enough to run on the third.

Cosmos 3 has also begun serving as the backbone for NVIDIA's downstream driving models. On August 4, 2026, NVIDIA open-released [NVIDIA Alpamayo 2 Super](https://aiwiki.ai/wiki/nvidia_alpamayo_2_super), a reasoning [vision-language-action model](https://aiwiki.ai/wiki/vision_language_action_model) for autonomous vehicles and robotaxis, stating that the model is "built on NVIDIA Cosmos 3 Super Reasoner" and post-trained with reinforcement learning, with the weights published on Hugging Face under the same OpenMDW-1.1 license as Cosmos 3.[^29] The "Cosmos 3 Super Reasoner" name refers to the Reasoner tower of Cosmos 3 Super: the Alpamayo 2 Super model card lists a 32-billion-parameter vision-language backbone, the size of the reasoner tower inside the 64-billion-parameter Cosmos 3 Super, paired with a 2.3-billion-parameter diffusion-based action decoder.[^29][^30]

## World action models and the DROID policies

On August 4, 2026, NVIDIA published a technical blog post, "Beyond VLAs: How World Action Models Reshape Robot Manipulation," that recast the Cosmos 3 policy checkpoints as [world action models](https://aiwiki.ai/wiki/world_action_model) (WAMs): robot policies post-trained on a video world model backbone rather than on the vision-language models that underpin most vision-language-action (VLA) systems.[^31] The argument is that a VLM backbone learns to describe a scene, not to predict how it evolves, so VLAs generalize well semantically but poorly to unseen physical motions; a policy built on a world model inherits a physics prior, which NVIDIA says translates into learning from more varied data, better behavior outside the training distribution, and fewer demonstrations when adapting to a new arm or gripper.[^31] The framing draws on the February 2026 NVIDIA research paper "World Action Models are Zero-shot Policies," which introduced DreamZero, a WAM built on a pretrained video diffusion backbone, and on NVIDIA researcher Jim Fan's "Robotics' End Game" talk.[^31][^32]

The post centers on **Cosmos3-Nano-Policy-DROID**, a 16-billion-parameter policy post-trained from Cosmos 3 Nano for the DROID platform, a Franka Panda arm with a Robotiq gripper (the setup behind the open DROID manipulation dataset).[^26][^31][^33] Its model card states the task plainly: given language instructions and visual observations from the DROID platform, generate robot action trajectories for manipulation and control.[^33] The weights have been on Hugging Face under the OpenMDW-1.1 license since the May 31, 2026 family launch, with only BF16 precision tested, so the blog repositions an existing checkpoint rather than announcing a new release.[^33] It is the workstation-class sibling of the 4B Cosmos3-Edge-Policy-DROID covered above: in NVIDIA's real-world DROID deployment the Nano policy runs beside the robot rather than on board, served on a single RTX PRO 6000 while the robot streams observations over the network and receives action chunks back.[^31]

Because post-training removes nothing from the omnimodel, the policy checkpoint keeps the full architecture: it can still reason and generate video, and when it outputs actions it can simultaneously output the video its cameras would see if those actions were executed.[^31] As evidence that the world-model prior rather than scale drives policy quality, the blog cites a controlled comparison from the Cosmos 3 technical report: two DROID policies trained with the same recipe, data and compute reached 28.1% RoboLab success starting from the base checkpoint versus 36.8% starting from an omni checkpoint trained on multi-domain action data.[^31] These are NVIDIA's own evaluations. The post closes by pointing developers to a Cosmos Labs livestream on WAMs, VLAs and hybrid robot foundation models scheduled for August 13, 2026, which NVIDIA Robotics promoted with the line that Cosmos 3 "natively represents action alongside video."[^31][^34]

## ELI5: what is Cosmos 3?

Imagine you want to teach a robot to pick things up, but you do not have the time to show it a million examples by hand. Cosmos 3 is like a very good imagination for machines. You can describe or show it a scene, it figures out what is going on and what should happen next, and then it draws a realistic video of it, or even writes down the exact movements a robot arm should make. Because it can dream up endless practice scenes, including weird ones that rarely happen in real life, robots and self-driving cars can learn faster and more safely. The smallest version, Cosmos 3 Edge, is light enough to live inside the robot itself, so it can think and act on the spot instead of asking a faraway computer for help. And because NVIDIA gave the models away openly, anyone can download them and build on them.

## See also

- [NVIDIA Cosmos](https://aiwiki.ai/wiki/nvidia_cosmos) (platform lineage and earlier models)
- [Physical AI](https://aiwiki.ai/wiki/physical_ai)
- [World model](https://aiwiki.ai/wiki/world_model)
- [Robot learning](https://aiwiki.ai/wiki/robot_learning)
- [Synthetic data](https://aiwiki.ai/wiki/synthetic_data)
- [Isaac GR00T](https://aiwiki.ai/wiki/isaac_gr00t)
- [NVIDIA Omniverse](https://aiwiki.ai/wiki/nvidia_omniverse)
- [Jetson Thor](https://aiwiki.ai/wiki/jetson_thor)
- [Diffusion model](https://aiwiki.ai/wiki/diffusion_model)
- [Vision language model](https://aiwiki.ai/wiki/vision_language_model)

## References

[^1]: NVIDIA Newsroom. "NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI." June 1, 2026. https://nvidianews.nvidia.com/news/nvidia-launches-cosmos-3-the-open-frontier-foundation-model-for-physical-ai

[^2]: NVIDIA Blog. "How Cosmos 3 Helps Physical AI Think Before It Acts." June 1, 2026. https://blogs.nvidia.com/blog/cosmos-3-physical-ai-open-world-foundation-model/

[^3]: NVIDIA Technical Blog. "Develop Physical AI Reasoning, World, and Action Models with NVIDIA Cosmos 3." June 1, 2026. https://developer.nvidia.com/blog/develop-physical-ai-reasoning-world-and-action-models-with-nvidia-cosmos-3/

[^4]: NVIDIA. "NVIDIA Cosmos: World Foundation Models Powering Physical AI" (product page). https://www.nvidia.com/en-us/ai/cosmos/

[^5]: Hugging Face. "Welcome NVIDIA Cosmos 3: The First Open Omni-model for Physical AI Reasoning and Action." 2026. https://huggingface.co/blog/nvidia/cosmos-3-for-physical-ai

[^6]: GitHub. "NVIDIA/Cosmos" repository. https://github.com/nvidia/cosmos

[^7]: NVIDIA Research, Cosmos Lab. "Cosmos 3: Omnimodal World Models for Physical AI" (technical report). 2026. https://research.nvidia.com/labs/cosmos-lab/cosmos3/technical-report.pdf

[^8]: Hugging Face. "nvidia/Cosmos3-Nano" model card. May 31, 2026. https://huggingface.co/nvidia/Cosmos3-Nano

[^9]: NVIDIA Blog. "NVIDIA GTC Taipei at COMPUTEX: Live Updates on What's Next in AI." June 1, 2026. https://blogs.nvidia.com/blog/nvidia-gtc-taipei-computex-2026-news/

[^10]: MarkTechPost. "NVIDIA Releases Cosmos 3: A Two-Tower Mixture-of-Transformers Foundation Model Unifying Physical Reasoning, World Generation, and Action Generation." June 3, 2026. https://www.marktechpost.com/2026/06/03/nvidia-releases-cosmos-3-a-two-tower-mixture-of-transformers-foundation-model-unifying-physical-reasoning-world-generation-and-action-generation/

[^11]: NVIDIA (Hugging Face blog). "Introducing Cosmos 3 Edge." July 20, 2026. https://huggingface.co/blog/nvidia/cosmos3edge

[^12]: NVIDIA Blog. "At SIGGRAPH, NVIDIA Advances Graphics and Simulation With Agentic and Physical AI." July 20, 2026. https://blogs.nvidia.com/blog/siggraph-news-2026/

[^13]: Hugging Face. "nvidia/Cosmos3-Edge" model card. July 20, 2026. https://huggingface.co/nvidia/Cosmos3-Edge

[^14]: Hugging Face. "nvidia/Cosmos3-Edge-Policy-DROID" model card. July 20, 2026. https://huggingface.co/nvidia/Cosmos3-Edge-Policy-DROID

[^15]: NVIDIA Blog. "NVIDIA Introduces New Jetson Thor Computers to Advance Mainstream Robotics and Edge AI." July 15, 2026. https://blogs.nvidia.com/blog/jetson-thor-robotics-edge-ai-agent/

[^17]: explainX. "NVIDIA SIGGRAPH 2026: Cosmos Edge and MCP." July 2026. https://explainx.ai/blog/nvidia-siggraph-2026-cosmos-edge-mcp-physical-ai-july-2026

[^18]: GitHub. "nv-tlabs/omni-dreams: NVIDIA Cosmos-Dreams (fka NVIDIA OmniDreams)" repository. 2026. https://github.com/nv-tlabs/omni-dreams

[^19]: Hugging Face. "nvidia/omni-dreams-models" model card. June 2, 2026. https://huggingface.co/nvidia/omni-dreams-models

[^20]: NVIDIA Research. "OmniDreams" project page. 2026. https://research.nvidia.com/labs/sil/projects/omnidreams-blog/

[^21]: Investing.com. "NVIDIA at SIGGRAPH 2026: AI Graphics and Simulation Push" (keynote transcript). July 20, 2026. https://www.investing.com/news/transcripts/nvidia-at-siggraph-2026-ai-graphics-and-simulation-push-93CH-4801795

[^22]: Vaidio. "Vaidio AI Vision Platform Brings Visual Reasoning to the Edge With NVIDIA Cosmos Reason 3 Edge, NVIDIA Jetson Thor and Vaidio Vision AI." July 2026. https://www.vaidio.ai/blog/bringing-vlm-reasoning-to-the-edge-with-nvidia-cosmos-reason-3-edge-nvidia-jetson-thor-and-vaidio-vision-ai

[^23]: Centific. "Centific Brings Real-Time Physical AI to the Edge with NVIDIA Cosmos 3 Edge." July 2026. https://www.centific.com/blog/centific-brings-real-time-physical-ai-to-the-edge-with-nvidia-cosmos-3-edge


[^24]: NVIDIA AI. "Introducing Cosmos 3 Edge: our open frontier world model built to run on-device." X post. July 20, 2026. https://x.com/NVIDIAAI/status/2079236204743053592

[^25]: Hugging Face. "nvidia/Cosmos3-Edge transformer/config.json." July 2026. https://huggingface.co/nvidia/Cosmos3-Edge/blob/main/transformer/config.json

[^26]: DROID. "A Large-Scale In-The-Wild Robot Manipulation Dataset." 2024. https://droid-dataset.github.io/

[^27]: OpenMDW. "OpenMDW License Version 1.1." June 2026. https://openmdw.ai/license/1-1/

[^28]: Linux Foundation. "Linux Foundation Releases OpenMDW 1.1; NVIDIA Adopts OpenMDW for Cosmos, Isaac GR00T, Ising, and Nemotron AI Model Families." May 28, 2026. https://www.linuxfoundation.org/press/linux-foundation-releases-openmdw-1.1-nvidia-adopts-openmdw-for-cosmos-isaac-gr00t-ising-and-nemotron-ai-model-families

[^29]: NVIDIA Blog. "NVIDIA Alpamayo 2 Super, the Frontier Open Model for Robotaxis and Autonomous Vehicles, Now Available for Commercial Use." August 4, 2026. https://blogs.nvidia.com/blog/alpamayo-2-super-open-model-now-available/

[^30]: Hugging Face. "nvidia/Alpamayo2-Super" model card. August 4, 2026. https://huggingface.co/nvidia/Alpamayo2-Super

[^31]: NVIDIA Technical Blog. "Beyond VLAs: How World Action Models Reshape Robot Manipulation." August 4, 2026. https://developer.nvidia.com/blog/beyond-vlas-how-world-action-models-reshape-robot-manipulation/

[^32]: arXiv. "World Action Models are Zero-shot Policies" (2602.15922). February 17, 2026. https://arxiv.org/abs/2602.15922

[^33]: Hugging Face. "nvidia/Cosmos3-Nano-Policy-DROID" model card. May 31, 2026. https://huggingface.co/nvidia/Cosmos3-Nano-Policy-DROID

[^34]: NVIDIA Robotics. "What comes after vision-language-action models in robotics?" X post. August 6, 2026. https://x.com/NVIDIARobotics/status/2085407590494634375

[^35]: NVIDIA Blog. "Into the Omniverse: How Open World Models Push the Frontier of Physical AI." August 6, 2026. https://blogs.nvidia.com/blog/open-world-models-physical-ai/

[^36]: Artificial Analysis. "Text to Image Leaderboard." Accessed August 8, 2026. https://artificialanalysis.ai/image/leaderboard/text-to-image

[^37]: Artificial Analysis. "Image to Video Leaderboard." Accessed August 8, 2026. https://artificialanalysis.ai/video/leaderboard/image-to-video

[^38]: NVIDIA et al. "Open Weights and American AI Leadership" (open letter). July 24, 2026. https://images.nvidia.com/pdf/Open-Weights-and-American-AI-Leadership.pdf

[^39]: NVIDIA Newsroom. "Japan's Robotics and Manufacturing Leaders Build on NVIDIA Cosmos to Advance Physical AI Frontier." July 15, 2026. https://nvidianews.nvidia.com/news/japans-robotics-and-manufacturing-leaders-build-on-nvidia-cosmos-to-advance-physical-ai-frontier

[^40]: PR Newswire (Afari Technology). "Afari Technology Unveils Its AI Plan." September 29, 2025. https://www.prnewswire.com/news-releases/afari-technology-unveils-its-ai-plan-302570297.html

[^41]: NVIDIA Technical Blog. "Post-Train NVIDIA Cosmos 3 Edge for On-Device Robot Control." August 19, 2026, modified August 20, 2026. https://developer.nvidia.com/blog/post-train-nvidia-cosmos-3-edge-for-on-device-robot-control/

[^42]: NVIDIA. "Cosmos3 DROID Action-Policy Post-Training" (Cosmos framework reproduction guide), revision c7e8d76b. Accessed August 20, 2026. https://github.com/NVIDIA/cosmos-framework/blob/c7e8d76b5da8aeae38cdac91c6cfd57185b2f6bc/docs/action_policy_droid_posttrain.md

[^43]: Hugging Face. "nvidia/Cosmos3-DROID" dataset card, revision 5c11a20a. Accessed August 20, 2026. https://huggingface.co/datasets/nvidia/Cosmos3-DROID/blob/5c11a20accb11497270a5247a7f1e66ad04c956c/README.md

[^44]: Khazatsky et al. "DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset." 2024, revised 2025. https://arxiv.org/abs/2403.12945

[^45]: Yang et al. "RoboLab: A High-Fidelity Simulation Benchmark for Analysis of Task Generalist Policies." Robotics: Science and Systems XXII, 2026. https://arxiv.org/abs/2604.09860

[^46]: Chi et al. "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion." The International Journal of Robotics Research, 2024. https://arxiv.org/abs/2303.04137

[^47]: NVIDIA Research. "NVlabs/RoboLab" source repository, revision 9db0aaf0. Accessed August 20, 2026. https://github.com/NVlabs/RoboLab/tree/9db0aaf09d9fe5d4f37b168320788258c7012463