An overview of the key announcements from the NVIDIA GTC 2026 conference – from new chips and systems to demonstrations and Jensen Huang's keynote.
SK Telecom has successfully completed the first phase of a national project to create a sovereign AI – a language model with 519 billion parameters.
AI: Events
Inside the Attention Mechanism: How PyTorch Tackles Problems Beyond the Standard Softmax
Technical context • Infrastructure
PyTorch has introduced a generalized attention mechanism, GDPA – an approach that allows the standard softmax operation in transformers to be replaced with any other function.
Yandex AI Studio has updated its file search tool, enabling AI agents to work with tables, audio, and video to find information in corporate knowledge bases.
AI: Events
Together AI Expands Model Fine-Tuning Capabilities: Now with Support for Tools, Reasoning, and Vision
Products
Together AI has updated its model fine-tuning service, adding support for tools, visual perception, and logical reasoning, and increasing training speeds by up to 6x.
AI: Events
Databricks Introduces New Embedding Model for Data Retrieval and Processing in AI Agents
Products
Databricks has released an embedding model to the public, designed to improve information retrieval accuracy in AI agents and corporate systems based on RAG architecture.
AI: Events
Mamba-3: Faster Than Transformers in Practice, Not Just on Paper
Technical context • Research
Mamba-3 has been released – an open-source language model that outpaces transformers in text generation speed and surpasses previous versions in quality.
AI: Events
On-Device Voice AI Agents: How PyTorch is Building a Unified Platform for Voice Tasks
Infrastructure
PyTorch has introduced an approach for running voice AI agents locally on devices – without the cloud, with support for multiple platforms and real-time tasks.
AI: Events
NVIDIA Nemotron 3 Super Now Available on Together AI: What This Means for Developers
Products
NVIDIA's Nemotron 3 Super launched on the Together AI platform on its official release day. The model is capable of multi-agent scenarios and can process up to one million tokens at once.