# Hugging Face Transformers v5.22.0 — Release & Upgrade Intelligence
- **Product:** Hugging Face Transformers (https://huggingface.co)
- **Version:** v5.22.0
- **Released:** Oct 29, 2026
- **Impact Score:** 9/10
- **Breaking Changes:** Source flagged
- **Canonical Analysis:** https://changelogs.us/transformers/updates/v5.22.0/

## Executive Summary (TL;DR)
Hugging Face Transformers v5.22.0 introduces the Mistral Large 4 1T-parameter model and adds support for FP8 adapter training. The release includes a breaking architectural change regarding Execution Provider (EP) logic separation.

## Breaking Changes & Deprecations
- Breaking changes * 🚨 Separate EP logic into its own file (#49679) by @3outeille

## New Features
- Support for Mistral Large 4 (1T parameter MoE model)
- FP8 quantization support for adapter training (LoRA/PEFT)
- Chunked expert dequantization and autograd skipping for FP8 backward passes

## Improvements
- Reduced peak memory in MoE layer backward passes from 28.9 GiB to 2.8 GiB at 8k tokens
- AMD CI image updates including onnxscript, bitsandbytes, and peft
- Deferred eager model imports in pipelines for improved startup performance

## Bug Fixes
- Fixed boolean padding in ClapFeatureExtractor
- Resolved string prompt iteration bug in Janus
- Fixed fp16 overflow in timesfm
- Preserved NumPy integer label dtypes in default data collators

## Upgrade Command
```bash
python -m pip install 'transformers==5.22.0'
```