Add enterprise GPU benchmarks and clarify FP8 model card

EXPLAINABILITY.md

@@ -5,12 +5,12 @@
 | Intended Application & Domain | World reasoning and generation for Physical AI. |
 | Model Type | Mixture-of-Transformers architecture with two towers. One is an autoregressive model for Physical AI reasoning; the other is a diffusion model for Physical AI generation. |
 | Intended Users | Physical AI developers, researchers, and practitioners building or evaluating autonomous vehicle, robotics, and world-generation workflows. |
-| Output | Images, videos, audio, and action commands. |
+| Output | Text, images, videos, audio, and action commands. |
 | Tools used to evaluate datasets to identify synthetic data and ensure data authenticity. | Dataset provenance analysis, metadata validation, watermark and artifact detection, embedding-based clustering, heuristic quality checks, and model-assisted data validation pipelines are used to identify synthetic content patterns, assess dataset authenticity, and improve data quality during dataset curation. |
 | Describe how the model works | Cosmos3 is an Omni world foundation model that generates texts, images, videos, audio, and action commands from combinations of text, images, videos, and action trajectory inputs. Input tokens from multiple modalities are packed into a shared sequence and processed by our mixture-of-transformer backbone with modality-specific output heads. |
 | Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of: | None. |
 | Technical Limitations | The model may not follow text, image, video, audio, or action trajectory inputs accurately in challenging cases, especially where the input contains complex scene composition, unusual camera motion, multiple interacting agents, low lighting, high motion blur, or fine-grained physical interactions. Generated outputs may contain temporal inconsistency, object morphing, inaccurate 3D structure, or implausible physical dynamics. Generated audio may not accurately render intelligible speech, or maintain strict temporal and semantic alignment with the visual context. |
 | Verified to have met prescribed NVIDIA quality standards | Yes. |
-| Performance Metrics | Video generation is measured using PAIBench-G, RBench, PhysicsIQ, and Artifical Analysis Image2Video benchmark. Image generation uses UniGenBench and Artifical Analysis Text2Image benchmark. For transfer evaluation, we use PAIBench-C and AVBench-C. Audio generation uses internal benchmarks. Action prediction uses metrics such as action MSE, Absolute Translation Error, Relative Translation Error, Relative Rotation Error, PSNR, and robotic task completion success rate. |
+| Performance Metrics | Video generation is measured using PAIBench-G, RBench, PhysicsIQ, and Artificial Analysis Image2Video benchmark. Image generation uses UniGenBench and Artificial Analysis Text2Image benchmark. For transfer evaluation, we use PAIBench-C and AVBench-C. Audio generation uses internal benchmarks. Action prediction uses metrics such as action MSE, Absolute Translation Error, Relative Translation Error, Relative Rotation Error, PSNR, and robotic task completion success rate. |
 | Potential Known Risks | This model can generate synthetic media and may produce content that is offensive, unsafe, misleading, indecent, or unsuitable for a target deployment. Users should implement robust safety guardrails — including content filtering, abuse monitoring, and access controls — to reduce the risk of harmful outputs. Users are responsible for ensuring that their use of the model complies with all applicable laws and regulations, and for regularly reviewing and updating their guardrails as risks evolve. |
 | Licensing | [OpenMDW1.1](https://openmdw.ai/)  |

README.md

@@ -91,13 +91,13 @@ Cosmos3 is an Omni-modal foundation model built on a Mixture-of-Transformers (Mo
     - Text: String
     - Image: jpg, png, jpeg, webp
     - Video (with or without audio): mp4
-    - Action: json (1D list)
+    - Action: JSON array shaped (T, D)
   - **Input Parameters**:
     - Text: One-dimensional (1D)
     - Image: Two-dimensional (2D)
     - Video: Three-dimensional (3D)
     - Audio: One-dimensional (1D)
-    - Action trajectory: One-dimensional (1D)
+    - Action trajectory: Two-dimensional (T, D)
   - **Other Properties Related to Input**:
     - For video inputs, we accept various resolutions, including 720p, 480p, and 256p.
     - When using input video with audio muxed into the video MP4 file, the audio should have 2 channels (stereo) and a 48 kHz sample rate.
@@ -116,17 +116,17 @@ Cosmos3 is an Omni-modal foundation model built on a Mixture-of-Transformers (Mo
     - Image: JPG
     - Video: MP4
     - Audio: Advanced Audio Coding (AAC) stream (muxed within the MP4)
-    - Action: 1D list (.json)
+    - Action: JSON object containing action data and shape metadata
     - Text: string
   - **Output Parameters**:
     - Image: Two-dimensional (2D)
     - Video: Three-dimensional (3D)
     - Audio: One-dimensional (1D)
-    - Action: One-dimensional (1D)
+    - Action: Two-dimensional (T, D)
     - Text: One-dimensional (1D)
   - **Other Properties Related to Output**:
     - The generated video is an MP4 file, with the resolution, frame rate, and duration specified in the input. The generated audio is encoded in AAC format, muxed into the video MP4 file with 2 channels (stereo) and a 48 kHz sample rate.
-    - Video generation supports durations from 5 to 400 frames, with 189 frames as the default generation duration.
+    - The technical paper lists video generation from 5 to 400 frames at 256p and 480p, and from 5 to 300 frames at 720p, at 10–30 fps. The examples below request 189 frames at 24 fps. Supported request sizes also depend on the inference backend.
     - The generated action is only supported for compatible embodiments, including general camera motion (9D), autonomous vehicle (9D), egocentric motion (57D), single Franka Panda arm with RobotiQ gripper (10D), dual Franka Panda arm with RobotiQ gripper (20D), Agibot (29D), UR (10D), Google robot (10D), WidowX 250 (10D), UMI (9D).
     - Audio: 48 kHz stereo AAC stream muxed into video mp4
     - Video: mp4 at the FPS specified in input
@@ -166,7 +166,7 @@ Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated sys
 
 **Runtime Engine(s):**
 
-- [PyTorch](https://github.com/nvidia/cosmos3)
+- [PyTorch / Cosmos Framework](https://github.com/nvidia/cosmos-framework)
 - [vLLM-Omni](https://github.com/vllm-project/vllm-omni)
 - [Hugging Face Diffusers](https://huggingface.co/docs/diffusers/en/index)
 - [SGLang](https://github.com/sgl-project/sglang)
@@ -183,6 +183,8 @@ Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated sys
 
 **Precision:** This repository provides the FP8-quantized Cosmos3-Nano checkpoint, including its quantization configuration and ModelOpt state. Components excluded from quantization retain their original precision.
 
+The Diffusers transformer configuration also specifies A16 activations for the first three and last three denoising iterations and for the Reasoner. Runtime support for this policy depends on the inference backend; FP8 does not imply that every component and operation runs in FP8.
+
 The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.
 
 ## Training, Testing, and Evaluation Datasets
@@ -264,6 +266,8 @@ Training datasets passed through multiple layers of automated and manual safegua
 
 Please see our [technical paper](https://research.nvidia.com/labs/cosmos-lab/cosmos3/technical-report.pdf) for detailed evaluations of the base model.
 
+The figures below report base-model quality evaluations. They do not establish the quality change caused by FP8 quantization of this checkpoint.
+
 ### Overall
 
 ![Overall benchmark results](https://huggingface.co/nvidia/Cosmos3-Nano/resolve/main/images/benchmark-overall.png)
@@ -296,16 +300,32 @@ All results use one GPU. Generator inference uses vLLM-Omni, Reasoner inference
 | H100 | 3.20 | 1.01 | 129.5 | 173.01 | 38.29 |
 | H200 | 3.34 | 1.00 | 128.3 | 202.05 | 38.97 |
 | H20 | 0.91 | 0.24 | 242.0 | 148.21 | 11.17 |
-| DGX Spark | — | — | 336.1 | 22.62 | — |
-| DGX Station | — | — | 85.4 | 255.99 | — |
+| DGX Spark | 0.39† | 0.11† | 336.1 | 22.62 | — |
+| DGX Station | 0.68† | 0.38† | 85.4 | 255.99 | — |
+| RTX PRO 5000 72GB Blackwell | 1.26† | 0.32† | — | — | — |
 | RTX PRO 6000 Blackwell Server Edition | 2.34 | 0.61 | 150.2 | 112.90 | 24.90 |
 
 - **Generator:** Image-to-video (i2v) only, at 480p and 720p. Throughput is calculated as **189 frames ÷ average end-to-end generation time in seconds**, following the existing model-card benchmark convention. This measures inference throughput for a generated sequence.
-- **Reasoner:** 50 input tokens, 100 output tokens, workload FPS 1, concurrency 1, and 50 requests. Both columns use the same workload. Output token throughput is the reported benchmark metric.
+- **Enterprise Generator (†):** Calculated from the measured times below using an assumed workload of 189 frames, as specified for this model-card update. The inspected enterprise report does not independently confirm the frame count or denoising-step count.
+- **Reasoner:** 50 input text tokens plus a 30-second 374×374 video sampled at 1 fps, 100 output tokens, concurrency 1, and 50 requests. Both columns use the same workload. Output token throughput is the reported benchmark metric, including request processing time; it is not decode-only token speed.
 - **Action Policy:** Policy Robot (DROID), 480p, with 32-action chunks. Action throughput is calculated as **32 ÷ average generation time in seconds**.
 - **GB300 Action Policy estimate (*):** Uses the single-GPU B300 result, assuming similar performance on GB300. This value has not been measured on GB300.
-- **Hardware:** H100 results use the 80GB HBM3 configuration; H200 results use the 141GB HBM3 configuration. RTX PRO 6000 results use the Blackwell Server Edition.
-- **Missing results:** A dash (—) indicates that benchmark results are unavailable. This applies to Generator and Action Policy measurements for DGX Spark and DGX Station.
+- **Hardware:** H100 results use the SXM 80GB HBM3 configuration; H200 results use the SXM 141GB HBM3e configuration. RTX PRO 6000 results use the Blackwell Server Edition.
+- **Missing results:** A dash (—) indicates that a verified value for that metric is not included here. The enterprise forward/inverse dynamics results are different workloads from DROID Action Policy.
+
+#### Enterprise Generator measurements
+
+The enterprise vLLM-Omni report (PBR 309485), published November 5, 2026 Pacific time, provides the following single-GPU FP8 image-to-video generation times. Values are the reported average generation times, without conversion to frames/s.
+
+| Platform (1 GPU) | Generator — 480p (seconds) ↓ | Generator — 720p (seconds) ↓ |
+| --- | ---: | ---: |
+| DGX Spark | 490.73 | 1669.62 |
+| DGX Station | 276.93 | 494.01 |
+| RTX PRO 5000 72GB Blackwell | 150.33 | 585.58 |
+
+DGX Spark and RTX PRO 5000 use CUDNN_ATTN; DGX Station uses Flash Attention. The report flags thermal slowdown on many DGX Spark tests and one DGX Station test. The frames/s table uses 189 frames for these workloads by assumption. The denoising-step count remains unverified, limiting direct comparisons with other results.
+
+Benchmark sources (NVIDIA network access required): [Generator, PBR 309431](https://prism.nvidia.com/ai-sim-platform/reports/40210), [datacenter Reasoner, PBR 308912](https://prism.nvidia.com/ai-sim-platform/reports/39455), [GB300 Reasoner, PBR 309589](https://prism.nvidia.com/ai-sim-platform/reports/40565), [enterprise Reasoner, PBR 308916](https://prism.nvidia.com/ai-sim-platform/reports/39470), and [enterprise Generator, PBR 309485](https://prism.nvidia.com/ai-sim-platform/reports/40954).
 
 ## Usage
 
@@ -369,7 +389,7 @@ The example inputs (`assets/`) live in this model repo. Download just this folde
 
 ```bash
 pip install -U "huggingface_hub[cli]"
-hf download nvidia/Cosmos3-Nano-FP8 assets/ --local-dir Cosmos3-Nano-FP8
+hf download nvidia/Cosmos3-Nano-FP8 --include "assets/*" --local-dir Cosmos3-Nano-FP8
 cd Cosmos3-Nano-FP8
 ```
 
@@ -1005,7 +1025,7 @@ curl -sS -L "http://localhost:30000/v1/videos/${job_id}/content" \
   -o cosmos3_t2v_output.mp4
 ```
 
-SGLang accepts Cosmos 3 request options including `max_sequence_length`, `flow_shift`, `extra_params.guardrails`, `extra_params.use_resolution_template`, and `extra_params.use_duration_template`. Video-to-video, video-with-sound, and action generation are not supported by SGLang yet.
+SGLang accepts Cosmos 3 request options including `max_sequence_length`, `flow_shift`, `extra_params.guardrails`, `extra_params.use_resolution_template`, and `extra_params.use_duration_template`. Modality support depends on the checkpoint and SGLang version. The current cookbook includes video-to-video, joint sound, and action workflows for compatible checkpoints; verify FP8 checkpoint support separately.
 
 For complete serving instructions and request examples, see the [Cosmos3 SGLang cookbook](https://docs.sglang.io/cookbook/diffusion/Cosmos/Cosmos3).
 
@@ -1021,7 +1041,7 @@ Cosmos3 outputs should not be treated as physically accurate simulation, reliabl
 
 **Test Hardware:** GB300, B300, B200, H100, H200, H20, DGX Spark, DGX Station, RTX PRO 6000 Blackwell Server Edition, RTX PRO 5000 72GB Blackwell
 
-See [Performance Benchmarking](#performance-benchmarking) for per-hardware results. Results not yet available are marked with a dash (—) and will be added as they come in.
+See [Performance Benchmarking](#performance-benchmarking) for per-hardware results, measurement units, and the meaning of missing values.
 
 ## Ethical Considerations