Hugging Face released transformers v5.20.1 on October 16, 2026, one day after v5.20.0. Beyond patch fixes, the release adds support for Beam, a new model from Reflection AI. Read the release notes.
The Beam model
Reflection AI describes Beam as a token-efficient, frontier-scale model focused on coding and agentic use cases, built for teams that want to run a powerful model on their own hardware, infrastructure or cloud and customize it to their needs. Transformers ships native documentation for the architecture.
Patch fixes
- Fixed a small bug in the auto mappings.
- Fixed HunYuan-VL image assignment during batched generation.
- Fixed DeepGEMM Mega MoE under expert parallelism.
from_pretrainedon torch 2.7 now requires torch 2.8 or newer for distributed checkpointing, and DtensorShardOperation works again on torch 2.6 and 2.7 — the release notes acknowledge that v5.20.0 broke loading models on older torch versions.
This report covers the October 16 release notes; the previous release is covered in our v5.20.0 report.