Hugging Face released transformers v5.19.0 on October 6, 2026. The release adds one new model architecture, several training and serving features, and a set of breaking changes. Read the release notes.
New model support
EmbeddingGemma2, a multimodal embedding model from Google built on the Gemma 4 architecture, embeds text, images, audio and video into a shared 768-dimensional space. It supports Matryoshka truncation to 512, 256 or 128 dimensions and allows disabling unused vision or audio towers.
Breaking changes to check
- MoE models whose routers compute logits now return them when
output_router_logits=True, following the Qwen3-MoE pattern. Owlv2ForObjectDetection.embed_image_querynow selects the query box with the highest objectness score, changing prior detection behavior.- The
paged|prefix for SDPA and flash attention is deprecated; their regular implementations support continuous batching directly, while eager attention still usespaged|eager.
New features
- Expert parallelism with a new
ep_dispatch_expertsplan, default for Qwen3 MoE and Mellum; Trainer supports expert parallel training. - Per-layer cache configuration for DynamicCache and StaticCache.
- Continuous batching support for XPU devices.
- Registered custom processing backends and a Trainer
loss_is_scaled_for_gaoption.
Notable fixes
Generation no longer mutates the user's quantized cache configuration, a DataCollatorForLanguageModeling seeding bug is fixed, and Qwen3-VL chunked prefill, WatermarkDetector counting, encoder repetition penalty and GPT-OSS SwiGLU config handling received corrections. CI now runs Python 3.11, and pandas is no longer a setup dependency.
Upgrade paths that depend on MoE outputs or the paged attention prefix should test against these changes first. This report covers the October 6 release notes; our verification guide explains how we track them.