Hugging Face AI Updates: June 25, 2026

1. NVIDIA NeMo AutoModel Brings 3.4x Fine-Tuning Speedup for MoE Models on Hugging Face

Hugging Face. NVIDIA published a post on the Hugging Face blog introducing NeMo AutoModel, a library built on Hugging Face Transformers v5 that accelerates Mixture-of-Experts model fine-tuning through Expert Parallelism, DeepEP fused dispatch, and TransformerEngine kernels. The approach achieves 3.4 to 3.7 times higher training throughput and 29 to 32 percent less GPU memory usage while remaining API-compatible with existing Transformers code via a single import change. The result makes full fine-tuning of frontier-scale MoE models such as the 550B Nemotron practical across multiple nodes without requiring developers to rewrite their training pipelines. Source

2. FFASR Leaderboard Launches to Benchmark Speech Recognition Under Real-World Conditions

Hugging Face. Researchers introduced the Far-Field ASR Leaderboard on Hugging Face, the first open community-driven benchmark designed to evaluate speech recognition models under realistic far-field acoustic conditions including room acoustics, background noise, and microphone distance. Existing ASR benchmarks use clean studio audio, which fails to predict performance in actual voice assistant deployments where degraded acoustics are the norm. The leaderboard aims to redirect model development toward robustness in real-world scenarios by making far-field performance measurable and comparable across models. Source