NVIDIA Cosmos Data Factory and Cosmos 3 updates

NVIDIA's Cosmos Data Factory blueprints and Cosmos 3 tools for generating, curating and evaluating physical-AI training data.

NVIDIA · Data, training, and fine-tuning platforms (datasets, labeling, RAG/vector DBs, fine-tuning-as-a-service) · tool

Published in Feed

This entry is marked for review. Check the original sources before relying on its details.

Overview

At GTC Berlin on October 21, 2026, NVIDIA released new Cosmos tools for training and evaluating physical-AI world models. Cosmos Data Factory labels, curates, augments and evaluates data with agentic skills, and ships five blueprints: autonomous-vehicle dataset generation, image attribute augmentation, event video generation, defect image generation and video data augmentation. Cosmos Framework tools, offered as agent skills, cover fine-tuning, reinforcement learning, distillation and quantization for open models including GR00T, Alpamayo, Qwen and Pi0.5. Updated Cosmos 3 NIM microservices claim up to 2x faster inference with no quality loss in validated configurations.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

  • Five Data Factory blueprints on GitHub, Microsoft Azure, Nebius and CoreWeave
  • Cosmos Curator with multimodal, multisensory input support
  • Cosmos 3 Nano simulation variants on Hugging Face for AV and bimanual-robot policy testing
  • Updated Cosmos 3 NIM microservices in the NGC catalog
  • Metropolis VSS 3.3 reported to monitor 40% more video streams on the same GPU
  • Reported top-five placement on six open-model leaderboards

Use cases

  • Generating synthetic defect images for circuit-board inspection (Siemens reported a 40% cut in calibration time at its Erlangen factory using a related approach)
  • Distilling a 32B reasoner into an 8B Nano model (MicroAGI)
  • Reducing repeated physical test drives for AV development (Porsche projects up to 90% fewer)

Pricing

Unknown; licensing terms for the models and tools are not stated in the announcement.

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released
21 Oct 2026
Record updated
27 Oct 2026
Record last checked
27 Oct 2026

Explore the Feed

Other entries filed under data, training, and fine-tuning platforms (datasets, labeling, rag/vector dbs, fine-tuning-as-a-service); inclusion does not imply equivalent features.

  • DigitalOcean Agent Data Services

    DigitalOcean's managed data layer for agents: ingestion flows, a catalog with vectors, and SQL or vector query, in private preview.

  • pplx-embed-v2-late

    Perplexity's late-interaction multimodal embedding models (0.6B and 9B) with 128-dim per-token vectors, open on Hugging Face.

  • EmbeddingGemma 2

    Open-weight 740M-parameter multimodal embedding model built on Gemma 4 that encodes text, images, audio, and video for on-device search.

Browse the Feed