Hugging Face AI Updates: July 18, 2026

1. NVIDIA NeMo Automodel Adds Distributed Fine-Tuning for Hugging Face Diffusers Models

Hugging Face. NVIDIA and Hugging Face detailed an integration between NVIDIA NeMo Automodel and the Diffusers library that enables production-grade distributed fine-tuning of diffusion models without checkpoint conversion or model rewrites. Users can point the training config at any Diffusers model ID on the Hub, and the system supports FSDP2, tensor, context, and pipeline parallelism configured via YAML to scale from a single GPU to hundreds, with both full fine-tuning and LoRA-style parameter-efficient options. The integration ships ready-to-use recipes for FLUX.1-dev, FLUX.2-dev, Wan 2.1 and 2.2, HunyuanVideo 1.5, and Qwen-Image, using flow-matching objectives, VAE latent pre-encoding, and multiresolution bucketed dataloading. NVIDIA reported FLUX.1-dev reaching roughly 35.51 images per second with full fine-tuning on 8 H100 GPUs, and the integration is released under the Apache 2.0 license. Source

2. Dharma-AI Says Specialized DharmaOCR Beats Newer Generalist OCR Models on Portuguese

Hugging Face. Dharma-AI published findings arguing that its specialized DharmaOCR model outperforms newer, more generalist OCR systems on Brazilian Portuguese tasks despite those competitors using more advanced architectures. On Portuguese-focused benchmarks, DharmaOCR scored 0.925 versus 0.798 for Mistral OCR4 and 0.7587 for Unlimited-OCR, gaps of 13 and roughly 16 points respectively. The team attributed the advantage to a two-stage training approach that concentrates parameters on Brazilian Portuguese vocabulary, morphology, and document structures through supervised fine-tuning, followed by Direct Preference Optimization to reduce text degeneration when models hit visual complexity. The post illustrated competitor failures such as misreading “Chico Buarque” as “Chico Barque,” pointing to limited exposure to Brazilian Portuguese proper nouns in multilingual models. Source