Hugging Face Transformers
Hugging Face Transformers library, backbone of the open AI stack.
Release history
Hugging Face Transformers Patch 5.20.2
Fixes a type-preservation bug in generate() affecting multimodal model inputs; Patch if running multimodal pipelines
Hugging Face Transformers 5.20.1
Adds support for Reflection AI's Beam model, a coding/agentic model. No impact unless loading that weight
Hugging Face Transformers 5.20.0
Transformers 5.20.0 adds native support for five multimodal and specialized models: Muse Spark Open, Ling 3.0 Flash, Kolibri1, MiniCPM-V 4.7, UnlimitedOcr, and UltraBERT
Hugging Face Transformers 5.19.0
Adds EmbeddingGemma 2, a multimodal embedding model on the Gemma 4 architecture. Relevant for anyone building retrieval or embedding pipelines on Transformers
Hugging Face Transformers 5.18.0
Adds Nemotron 3 Diarization model for streaming speaker diarization
Hugging Face Transformers 5.17.0
Adds HYV4, a 780B MoE model activating 49B parameters per token
Hugging Face Transformers 5.16.1
Adds GLM model support
Hugging Face Transformers Release 5.16.0
Adds Qwen4-Exp model with GatedResidual, Qwen Sparse Attention, and Per-Layer Embedding components
Hugging Face Transformers Patch release 5.15.1
Fixes DFlash and MTP candidate generator issues, resolves image processing on accelerators with Lanczos filter
Hugging Face Transformers Release 5.15.0
Adds Meta Muse Glimmer model support for agentic multimodal workflows
Hugging Face Transformers Patch release 5.14.1
Fixes assisted generation issue with EncoderDecoderCache models
Hugging Face Transformers 5.14.0
Adds Inkling multimodal model: 975B total parameters, 41B active, accepts text, image, and audio inputs
Hugging Face Transformers Patch 5.13.1
Compatibility patch for vLLM 0.25.0
Hugging Face Transformers 5.13.0
Adds KimiK 2.5, 2.6, 2.7 architecture support
Hugging Face Transformers Patch 5.12.1
Raises PEFT lower bound; Fixes mistral tokenizer resolution when mistral-common is installed
Hugging Face Transformers Patch 5.10.3
Fixes for vLLM integration, token ID handling, InternVL models, and processor offsets in Transformers 5.10.3
Hugging Face Transformers 5.12.0
Adds MiniMax-M3-VL vision-language model support
Hugging Face Transformers Patch 5.10.2
Fixes model conversion bug for CLIP-related models including SAM3. Update if those weights are in use
Hugging Face Transformers 5.10.1
5.10.0 was yanked; Upgrade directly to 5.10.1
Hugging Face Transformers 5.9.0
Adds Cohere2Moe Command A+ MoE model with hybrid attention
Hugging Face Transformers Patch 5.8.1
Fixes Deepseek V4 integration. Update if using that model
Hugging Face Transformers 5.8.0
Transformers 5.8.0 adds DeepSeek-V4, Gemma4 Assistant for speculative decoding, GraniteSpeechPlus speech-to-text, Granite4Vision for document extraction, and EXAONE-4.5 vision-language model