Hugging Face · 2026-11-21 · notable
Transformers v5.24.0 — Omnilingual ASR, Inkling 2 Mini and one cache name
In short
Transformers v5.24.0 adds Meta's Omnilingual ASR for 1,600+ languages, Inkling 2 Mini, NVIDIA's looped Dory, Apertus 1.5, HY4-VL, Sarvam MoE and EXAONE. past_key_values and cache_params are now renamed to cache.
Hugging Face · 2026-11-21 · 2 sources checked

New speech, vision and looped models land in Hugging Face's library, and the KV-cache argument gets one name.
What is it?
Transformers v5.24.0 adds support for Meta's Omnilingual ASR speech models (more than 1,600 languages), Thinking Machines' Inkling 2 Mini, NVIDIA's Dory looped language model, the Apertus 1.5 text-image-audio model, HY4-VL, the WavTokenizer audio codec, Translated's Lara Mini and the architectures behind Sarvam AI's 105B and 30B mixture-of-experts models. LG's EXAONE 3.0, 3.5 and Deep models now load through the Llama classes.
How does it work?
The library now calls the generation cache argument `cache` everywhere, replacing `past_key_values` and `cache_params`; the old names keep working with a warning until v5.40.0 for forward calls and v6.0.0 for generate(). Static cache now works with assisted decoding, continuous batching runs under inference mode, and generative multimodal models share common output classes.
Why does it matter?
Code that passes `past_key_values` by hand should move to `cache` before the old name is removed. Upgraders also need peft 0.22.0 or newer, remote code must import masks from `transformers.masking_utils` because `modeling_attn_mask_utils.py` is gone, and loading pickled `.pt2` exports now requires trust_remote_code=True.
Who is it for?
ML engineers using Hugging Face models
Try it
pip install -U transformers==5.24.0