# Hugging Face Transformers v5.21 — Release & Upgrade Intelligence
- **Product:** Hugging Face Transformers (https://huggingface.co)
- **Version:** v5.21
- **Released:** Oct 26, 2026
- **Impact Score:** 7/10
- **Breaking Changes:** Source flagged
- **Canonical Analysis:** https://changelogs.us/transformers/updates/v5.21/

## Executive Summary (TL;DR)
Hugging Face Transformers v5.21. Source highlight: MiniCPM-OCR is a document-understanding model from OpenBMB's MiniCPM-V family. It keeps the MiniCPM-V 4.6 architecture, which pairs a SigLIP vision encoder and a window-attention merger with a Qwen3.5 language model. It is specialized for reading text-dense images such as scanned pages, screenshots,

## Breaking Changes & Deprecations
- Breaking changes Fixed Gemma4 processor imports (the Gemma4, Gemma4 unified and EmbeddingGemma2 processors) so `from transformers import *` no longer fails in environments that have `PIL` but not `torch`; no user action is needed.
- Fix Windows path separators breaking relative imports of Hub custom-generate modules (#48951) by @dltsum in [#48951]

## New Features
- Feat test (#49920) by @guarin in #49920
- Add Flint model support. (#49824) by @lovishmadaan in #49824
- Add UltraVBERT (#49685) by @athrael-soju in #49685
- Add GradiumPhononTTS (#49791) by @eustlb in #49791
- add PocketTTS (#49589) by @eustlb in #49589
- add PocketTTSAudioTokenizer (#49588) by @eustlb in #49588

## Improvements
- [CI] Improve run-slow PR comments and remove CI recap (#49782)

## Bug Fixes
- :rotating_light: Fix gemma4 cross imports resulting in import failing without `torch` but with `PIL` (#49748)
- [FA] Fix FA with static cache (#49998) by @Cyrilvallez in [#49998]
- Fix Qwen4-Exp indexer position cache lifecycle (#49840) by @taking-lying-flat in [#49840]
- Fix BLOOM crash with static cache / compiled generate (#49481) by @kaixuanliu in [#49481]
- [Fetcher] Fix missing test deps for dynamically-resolved image processors (#49826) by @ydshieh in [#49826]
- fix(vision): Route the DETR conversion scripts through from_pretrained (#49830) by @harshaljanjani in [#49830]