Open Source AI Week - How Developers and Contributors Are Advancing AI Innovation

As Open Source AI Week comes to a close, we're celebrating the innovation, collaboration and community driving open-source AI forward. Catch up on the highlights and stay tuned for more announcements coming next week at NVIDIA GTC Washington, D.C.

Wrapping Up a Week of Open-Source Momentum From the stages of the PyTorch Conference to workshops across Open Source AI Week, this week spotlighted the creativity and progress defining the future of open AI.

Here are some highlights from the event:

Honoring open-source contributions: Jonathan Dekhtiar, senior deep learning framework engineer at NVIDIA, received the PyTorch Contributor Award for his key role in designing the release mechanisms and packaging solutions for Python software and libraries that enable GPU-accelerated computing.

CEO of Modular visits the NVIDIA booth: Chris Lattner, CEO of Modular and founder and chief architect of the open-source LLVM Compiler Infrastructure project, picks up the NVIDIA DGX Spark.

Seven questions with founding researcher at fast.ai: Jeremy Howard, founding researcher at fast.ai and advocate for accessible deep learning, shares his insights on the future of open-source AI.

In his keynote at the PyTorch Conference, Howard also highlighted the growing strength of open-source communities, recognizing NVIDIA for its leadership in advancing openly available, high-performing AI models.

The one company, actually, that has stood out, head and shoulders above the others, and that is two, he said. One is Meta the creators of PyTorch. The other is NVIDIA, who, just in recent months, has created some of the world's best models - and they are open source, and they are openly licensed.

vLLM Adds Upstream Support for NVIDIA Nemotron Models Open-source innovation is accelerating. NVIDIA and the vLLM team are partnering to add vLLM upstream support for NVIDIA Nemotron models, transforming open large language model serving with lightning-fast performance, efficient scaling and simplified deployment across NVIDIA GPUs.

vLLM's optimized inference engine empowers developers to run Nemotron models like the new Nemotron Nano 2 - a highly efficient small language reasoning model with a hybrid Transformer-Mamba architecture and a configurable thinking budget.

Learn more about how vLLM is accelerating open model innovation.

NVIDIA Expands Open Access to Nemotron RAG Models NVIDIA is making eight NVIDIA Nemotron RAG models openly available on Hugging Face, expanding access beyond research to include the full suite of commercial models.

This release gives developers a wider range of tools to build retrieval-augmented generation (RAG) systems, improve search and ranking accuracy, and extract structured data from complex documents.

The newly released models include Llama-Embed-Nemotron-8B, which provides multilingual text embeddings built on Llama 3.1, and Omni-Embed-Nemotron-3B, which supports cross-modal retrieval for text, images, audio and video.

Developers can also access six production-grade models for text embedding, reranking and PDF data extraction - key components for real-world retrieval and document intelligence applications.

With these open-source models, developers, researchers and organizations can more easily integrate and experiment with RAG-based systems.

Developers can get started with Nemotron RAG on Hugging Face.

Building and Training AI Models With the Latest Open Datasets NVIDIA is expanding access to high-quality open datasets that help developers overcome the challenges of large-scale data collection and focus on building advanced AI systems.

The latest release includes a collection of Nemotron-Personas datasets for Sovereign AI. Each dataset is fully synthetic and grounded in real-world demographic, geographic and cultural data - with no personally identifiable information. The growing collection, which features personas from the U.S., Japan and India, enables model builders to design AI agents and systems that reflect the linguistic, social and contextual nuance of the nations they serve.

NVIDIA earlier this year released the NVIDIA Physical AI Open Datasets on Hugging Face, featuring more than 7 million robotics trajectories and 1,000 OpenUSD SimReady assets. Downloaded more than 6 million times, the datasets combines real world and synthetic data from the NVIDIA Cosmos, Isaac, DRIVE and Metropolis platforms to kickstart physical AI development.

NVIDIA Inception Startups Highlight AI Innovation

At the PyTorch Conference's Startup Showcase, 11 startups - including members from the NVIDIA Inception program - are sharing their work developing practical AI applications and connecting with investors, potential customers and peers.

Runhouse, an AI infrastructure startup optimizing model deployment and orchestration, was crowned the 2025 PyTorch Startup Showcase Award Winner. The Community Choice Award was presented to CuraVoice, with CEO Sakhi Patel, CTO Shrey Modi, and advisor Rahul Vishwakarma accepting the award on behalf of the team.

CuraVoice provides an AI-powered voice simulation platform - powered by NVIDIA Riva for speech recognition and text-to-speech, and NVIDIA NeMo for conversational AI models - for healthcare students and professionals, offering interactive exercises and adaptive feedback to improve patient communication skills.

Shrey Modi, CTO of CuraVoice, accepts the PyTorch Startup Showcase Community Choice Award. In addition to CuraVoice, other Inception members, including Backfield AI, Graphsignal, Okahu AI, Snapshot AI and XOR, were featured participants in the Startup Showcase.

Snapshot AI delivers actionable, real-time insights to engineering teams using recursive retrieval-augmented generation (RAG), transformers and multimodal AI. The company's platform taps into the NVIDIA CUDA Toolkit to deliver high-performance analysis

LINK:	https://blogs.nvidia.com/blog/open-source-ai-week/...
	See more stories from nvidia

More from Nvidia