Transformers v5.0.0rc-0 marks a major milestone in the evolution of the popular machine learning library, five years after the release of v4.0.0rc-1. Today, Transformers is installed over 3 million times daily via pip—up from just 20,000 per day in v4—and has surpassed 1.2 billion total installations. The ecosystem has grown dramatically, expanding from 40 model architectures in v4 to over 400 today, with more than 750,000 model checkpoints contributed to the Hugging Face Hub, compared to around 1,000 at the time of v4. This growth reflects the broader adoption of AI and the increasing reliance on standardized, open-source tools. To remain relevant, Transformers has undergone a significant transformation focused on simplicity, training, inference, and production readiness. Simplicity is now central to the library’s design. The team has prioritized clean, readable code that makes model definitions transparent and easier to understand. This promotes standardization, generality, and broader community trust. A modular architecture has been introduced to streamline contributions and reduce maintenance overhead. For example, the AttentionInterface abstraction centralizes attention mechanisms, allowing implementations like FlashAttention, FlexAttention, and SDPA to be managed independently from core model files. The team has also developed machine learning-powered tooling to identify similarities between new models and existing architectures, enabling automated draft pull requests for integration. This reduces manual effort and ensures consistency across model definitions. Codebase simplification has been a major focus. Modeling files now contain only essential components for forward and backward passes, while common utilities are abstracted away. Tokenization has been streamlined by retiring the distinction between “Fast” and “Slow” tokenizers. The library now uses tokenizers as the default backend, with optional support for SentencePiece and MistralCommon-based tokenizers. Image processors will now only be available in their fast variants, powered by torchvision. In a strategic shift, Transformers is focusing exclusively on PyTorch as its primary backend, sunsetting native Flax and TensorFlow support. However, partnerships with JAX-based tools ensure continued interoperability. Training capabilities have been significantly enhanced. Support for large-scale pre-training has been strengthened with improved model initialization, compatibility with parallelism frameworks, and optimized kernels for forward and backward passes. Transformers now integrates with tools like torchtitan, megatron, and nanotron. Fine-tuning and post-training workflows are better supported across ecosystems. Transformers now works seamlessly with tools like Unsloth, Axolotl, LlamaFactory, TRL, and MaxText, enabling agentic use cases through platforms like OpenEnv and the Prime Environment Hub. Inference has seen major improvements. New APIs simplify batched evaluation and support specialized kernels. Transformers now automatically uses optimized kernels when hardware and software allow. It remains compatible with leading inference engines like vLLM, SGLang, TensorRT, ONNXRuntime, llama.cpp, and MLX. This interoperability allows models added to Transformers to be instantly available in these engines, leveraging their optimizations for dynamic batching, low-latency inference, and efficient deployment. Local and on-device inference is being expanded through collaboration with executorch, with support for multimodal models and efficient quantization. Quantization is now a first-class feature in v5, with full support for 8-bit and 4-bit models. The weight loading system has been restructured to make quantization seamless and reliable. This enables better integration with tools like bitsandbytes and TorchAO, supporting advanced features like tensor parallelism and mixture-of-experts. The overarching theme of v5 is interoperability. The library now serves as a unified foundation across the AI stack—train with Unsloth or Axolotl, deploy with vLLM or SGLang, export to GGUF for llama.cpp or MLX, and run locally with executorch. This release is the result of five years of community-driven innovation. With v5.0.0rc-0 now available, the team is eager to receive feedback and continues to refine the library based on user input. The release notes provide full technical details, and the community is invited to share thoughts and report issues on GitHub.