Hugging Face AI Updates: July 9, 2026

1. NVIDIA Releases Open Nemotron Data for Building AI Agents

Hugging Face. NVIDIA published a post on the Hugging Face blog detailing its Nemotron open-data initiative, releasing over 10 trillion pre-training tokens and millions of post-training samples for developing AI agents. The release includes region-specific synthetic personas and an interactive Nemotron Post-Training v3 Prompt Atlas, using synthetic data to preserve organizational confidentiality. Source

2. Transformers’ vLLM Backend Reaches Native Inference Speed

Hugging Face. The Transformers modeling backend integrated with vLLM now reaches native vLLM inference speed for compatible models without hand-written optimization code. The update uses torch.fx for static analysis to automatically fuse operations and apply parallelization at runtime, letting models such as Qwen3 match or exceed custom vLLM implementations from a single trainable codebase. Source