Hugging Face AI Updates: July 5, 2026

1. Transformers v5.13.0 Adds Kimi K2.5, MiMo-V2-Flash, and a Wave of New Model Architectures

Hugging Face. The Transformers library shipped v5.13.0 with support for a large batch of new model families, including the Kimi K2.5, 2.6, and 2.7 multimodal agentic models, the MiMo-V2-Flash mixture-of-experts model with a 256K context window, and Zyphra’s ZAYA MoE. The release also introduces speech models such as Nemotron 3.5 ASR with configurable streaming chunk sizes and Qwen3 ASR with automatic language detection, alongside vision models VideoPrism, RADIO, and MiniCPM3. The additions broaden coverage across text, speech, and video understanding in a single update. Source

2. Transformers v5.13.0 Unifies Export Tooling and Standardizes Modeling Internals

Hugging Face. The same v5.13.0 release adds a unified HfExporter path targeting PyTorch, ONNX, and ExecuTorch, aiming to make model export more consistent across backends. It also lands breaking modeling refactors that standardize layer declarations and attention handling for export compatibility, which affect internal APIs. Maintainers flagged that Gemma 3 and Gemma 4 attention masking fixes may change result reproducibility for existing pipelines. Source