OffNet Newsroom

Daily topic roundup

AI / ML

Saturday, August 29, 2026 · 2 stories, curated & summarized — click any story for the source.

Share this roundup on LinkedIn Full edition
AWS What's New awsdatabase ↺ since 08-28

NVIDIA Cosmos3 models now on SageMaker JumpStart for physical AI

NVIDIA has released the Cosmos3-Edge, Cosmos3-Nano, and Cosmos3-Super foundation models on Amazon SageMaker JumpStart. These open, frontier omnimodal world models enable developers to build robots and autonomous systems that perceive, reason, and act in the physical world. Cosmos3-Edge specifically targets on-device control with a 4B-parameter model optimized for real-time visual reasoning at 15 Hz.

  • SageMaker JumpStart now hosts NVIDIA's full Cosmos3 family for physical AI development.
  • Cosmos3-Edge is a 4B-parameter model designed for real-time robot control and edge inference.
  • Models support vision AI tasks requiring perception, reasoning, planning, and action.
  • Cosmos3-Edge operates at 640x360 resolution and generates 32 actions per inference at 15 Hz.
BY THE NUMBERSCosmos3-Edge Inference Speed15 HzReal-time visual reasoning4B-parameter model for edge robot control
Read at source →
Hugging Face Blog llmaiml ↺ since 08-28

Hugging Face: Training Multi-Vector Embedding Models with Sentence Transformers

Hugging Face released guidance on training and fine-tuning multi-vector embedding models using the Sentence Transformers library. The approach focuses on encoder architectures capable of generating multiple vector representations per input, distinct from traditional single-vector embeddings. This enables more granular retrieval and similarity matching capabilities in downstream applications.

  • Use Sentence Transformers for streamlined multi-vector model training
  • Multi-vector embeddings allow for finer-grained semantic retrieval
  • Follow Hugging Face's specific fine-tuning recipes for best results
  • Suitable for advanced RAG pipelines requiring dense vector matching
Read at source →
↑