ContainersNVIDIADeepSeek-R1NVIDIALlama-3.1-Nemotron-70B-InstructNVIDIAPyTorch
CollectionsNVIDIAOmniverse Kit (FB)NVIDIADeepStream SDKNVIDIAOmniverse Kit App Streaming
ModelsNVIDIAStyleGAN3 pretrained modelsNVIDIAPeopleNetNVIDIATrafficCamNet
ResourcesNVIDIARiva Skills Quick Start
Helm ChartsNVIDIAGPU OperatorNVIDIANVIDIA NIM Operator
NVIDIA
NVIDIA
Cosmos-Embed1
Model
NVIDIA
NVIDIA
Cosmos-Embed1

Cosmos-Embed1 is a joint video-text embedder tailored for physical AI.

This model is backed by NVIDIA's Plus Plus (++) Promise
to learn more about the quality of the datasets used to train this model.
FieldResponse
Intended Task/Domain:Joint video-text embedding for physical AI applications (robotics, autonomous vehicles (AV), search, general video understanding) and video anomaly detection and classification.
Model Type:Transformer
Intended Users:Physical AI developers working on robotics, autonomous vehicles (AV), search, video understanding, and video anomaly detection tasks.
Output:L2-normalized text and video embedding vectors (256 dimensions for 224p inputs; 768 dimensions for 336p/448p inputs).
Describe how the model works:Video frames are individually processed by a ViT backbone, temporally augmented, and compressed via QFormer cross-attention into a single video embedding. Text is processed by the QFormer self-attention branch into a text embedding. Both embeddings are aligned via contrastive learning.
Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of:Not Applicable
Technical Limitations & Mitigation:Due to the training datasets being predominantly composed of short action-focused English captions, different text prompts may not be properly aligned to video data, producing suboptimal matching. The model has been optimized for 8 frames at 1–2 FPS; significantly different frame rates or video lengths may degrade performance. Fine-tuning on domain-specific data can mitigate these limitations.
Verified to have met prescribed NVIDIA quality standards:Yes
Performance Metrics:Retrieval metrics (top-K hit rate, mean reciprocal rank) and classification metrics (precision, recall, F1)
Potential Known Risks:The model may produce suboptimal embeddings to unsafe text prompts.
Licensing:GOVERNING TERMS: Use of this model is governed by the NVIDIA Open Model License. Additional Information: Apache 2.0 and MIT.