Hugging Face Transformers v5.20.1 adds Reflection AI's Beam model and fixes loading on older torch

Hugging Face's Transformers library v5.20.1 adds support for Beam, a new model from Reflection AI, and fixes a regression that broke model loading on older torch versions. The release notes call it a "special surprise" and link to a new beam documentation page.

What Hugging Face says about Beam

The release notes describe Beam as "a token-efficient, frontier-scale model focused on coding and agentic use cases." They say it is built for teams that want to run a powerful model on their own hardware, infrastructure, or cloud, and customize it to their needs.

These are descriptions from the release notes, not independently verified results. The notes do not include:

• Parameter count: not disclosed
• Context window: not disclosed
• Pricing per 1M tokens: not disclosed
• Benchmark scores: not disclosed
• Training cutoff date: not disclosed
• License and weight availability: not stated in the release notes

The notes also do not say whether Beam weights are publicly downloadable. The release adds the model's architecture and documentation to the library. A companion fix for the auto mappings (#49664, by @yonigozlan) is listed alongside the Beam addition.

Patch fixes

Hugging Face acknowledges that it "broke loading models on older torch versions" and says the issue should now be fixed. The listed fixes are:

• Older torch loading (#49652): Fixes from_pretrained on torch 2.7. Distributed checkpointing now requires torch>=2.8.
• DtensorShardOperation (#49657): Now works on torch 2.6 and 2.7.
• DeepGEMM Mega MoE (#49428): Fixes behavior under expert parallelism.
• HunYuan-VL (#49557): Fixes image assignment during batched generation.
• Auto mappings (#49664): Fixes a small bug.

The full diff is available at v5.20.0...v5.20.1 on GitHub.

What this means

The Beam addition is the main item. Reflection AI is positioning Beam as a model teams can self-host and customize, and the release notes say so directly. Day-one Transformers support lowers the integration barrier for anyone already using the library for inference or fine-tuning.

The notes leave open the details that decide whether Beam is competitive: size, context length, license, and measured coding performance. Treat the positioning as a company claim until Reflection AI publishes those figures or third-party evaluations appear.

The patch half of the release has a practical warning. Teams pinned to torch 2.6 or 2.7 hit loading failures in the prior release. Anyone on those versions should upgrade to v5.20.1 rather than v5.20.0. Distributed checkpointing now requires torch>=2.8, so multi-node setups on older torch need to plan an upgrade.