Transformers: Model framework for NLP and vision
Unified API for pre-trained transformer models across frameworks.
Learn more about transformers
Transformers is a Python library that centralizes model definitions for transformer architectures and related deep learning models. It implements model classes compatible with PyTorch and other frameworks, allowing the same model definition to work across different training and inference backends. The library includes pre-trained model weights, tokenizers, and feature extractors for various tasks including natural language processing, computer vision, speech recognition, and multimodal applications. Models defined in Transformers integrate with training frameworks like DeepSpeed, FSDP, and PyTorch-Lightning, as well as inference engines like vLLM.
Framework-Agnostic Models
Model definitions are decoupled from specific frameworks, enabling the same model to be used across PyTorch, TensorFlow, and JAX backends. Training and inference tools build on a shared model specification without reimplementation.
Unified Model Hub
Direct integration with a centralized repository containing thousands of pre-trained checkpoints. Single function call loads models with weights, configurations, and tokenizers, eliminating manual download and setup.
Multi-Modal Architecture Support
Provides unified interfaces for text, vision, audio, and multi-modal models. Process different data types with consistent APIs, enabling straightforward development of applications combining multiple modalities.
from transformers import pipeline
classifier = pipeline("sentiment-analysis")
result = classifier("I love using transformers!")
print(result)Patch release fixing an optional dependency parsing error with optax that broke poetry installations.
- –Update poetry lockfiles if you encountered optax dependency resolution errors in v4.57.0.
- –Verify FSDP workflows with pre-quantized BitsAndBytes checkpoints now load correctly after internal fix.
Adds five new model architectures (Qwen3-Next, Qwen3-VL, VaultGemma, LongCat-Flash, FlexOlmo) with no breaking changes reported.
- –Integrate Qwen3-Next for 80B-parameter MoE with 1:50 sparsity, delivering 10x throughput over 32K context vs. Qwen3-32B.
- –Adopt VaultGemma (1B params) for differential-privacy workloads with ε≤2.0 sequence-level DP guarantee via DP-SGD training.
Patch release fixing processor multi-processing, RoPE in Jetmoe, config getter regression, and Emu3 dtype parsing.
- –Update if you use processors with multi-processing or rely on Jetmoe RoPE correctness.
- –Fixes config attribute getter regression and Emu3 dtype edge case; no migration required.
See how people are using transformers
Related Repositories
Discover similar tools and frameworks used by developers
ComfyUI
Visual graph-based diffusion model workflow builder.
tiktoken
Fast BPE tokenizer for OpenAI language models.
ComfyUI-Manager
Graphical package manager for ComfyUI custom nodes.
goose
LLM-powered agent automating local software engineering workflows.
ollama
Go-based CLI for local LLM inference and management.