Hugging Face released transformers v5.20.0 on October 15, 2026. The release adds five new model architectures alongside fixes. Read the release notes.
New model support
- Muse Spark Open (Meta): a multimodal mixture-of-experts model that takes text, images, audio and video as input and generates text. It has 816B total parameters (42B active per token) and a context window of up to 1,048,576 tokens shared across modalities, and is trained for agentic coding tasks such as multi-file editing, debugging and tool use. Checkpoints ship in BF16 and NVFP4 formats.
- Ling 3.0 Flash (InclusionAI): a hybrid MoE language model whose decoder groups five Kimi Delta Attention layers with one Multi-head Latent Attention layer. It is natively supported as
BailingHybridForCausalLM, loading withouttrust_remote_code=True. - Kolibri 1 (Aleph Alpha): a German-English MoE reasoning model; each of its 50 layers routes every token to 6 of 384 experts with sigmoid routing plus one shared expert, alternating sliding-window and full-attention layers.
- MiniCPM-V 4.7 (OpenBMB): pairs a SigLIP vision encoder with a Qwen3.5 backbone and adds canvas M-RoPE, assigning each visual token a (temporal, height, width) position on a 2-D canvas so image slices keep spatial relationships and video frames keep temporal order.
- UnlimitedOcr (Baidu): a single 3B-parameter OCR model with a 32,768-token context, building on DeepSeek OCR.
This report covers the October 15 release notes; the previous release is covered in our v5.19.0 report. Our verification guide describes how we track library releases.