transformers

huggingface/transformers

Hugging Face Standard Library for Pretrained Multimodal Modelsready

Transformers is Hugging Face's core Python library that unifies the definition and usage of text, image, audio, video, and multimodal models. It is used by researchers, ML engineers, and application developers to load pretrained models and integrate them into training, inference, and serving pipelines. Releases of this package directly impact new model support, API compatibility, and integration stability with PyTorch/JAX/TF and various inference and training ecosystems.

machine-learningnlpcomputer-visionmultimodalpretrained-modelshuggingface

Choose version

5.16.01mo ago

huggingface/transformers Release: v5.16.0

Original

Key point

Added Qwen4-Exp and GraniteSpeech5 models, strengthening hybrid architecture and speech recognition capabilities.

Details

Added the Qwen4-Exp model: built on Qwen3.5's hybrid text and multimodal architecture, it introduces three core components: GatedResidual (GR), Qwen Sparse Attention (QSA), and Per-Layer Embedding (PLE). QSA integrates linear and sparse attention, significantly improving inference efficiency for long-context workloads.
Added the GraniteSpeech5 model: a lightweight conformer encoder with approximately 470M parameters, an automatic speech recognition (ASR) model trained on BPE targets using CTC (Connectionist Temporal Classification). Its encoder-only structure supports fast transcription.

transformers releases | AI Briefing