Transformers v5.20.1 adds Reflection AI's Beam model
A coding-focused model from Reflection AI lands amid a run of patch fixes
The facts
- Transformers v5.20.1 adds support for Beam, a new model from Reflection AI described as token-efficient and frontier-scale, built for coding and agentic use cases
- The release fixes a regression that had broken model loading on older PyTorch versions
- Other fixes cover HunYuan-VL image assignment in batched generation, DeepGEMM under expert parallelism for the Mega MoE model, and DtensorShardOperation compatibility with PyTorch 2.6 and 2.7
- from_pretrained on PyTorch 2.7 now requires torch 2.8 or later for distributed checkpointing
Hugging Face has shipped Transformers v5.20.1, adding support for Beam, a new model from Reflection AI. The release notes describe Beam as a token-efficient, frontier-scale model built for coding and agentic use cases. It is aimed at teams that want to run a powerful model on their own hardware, infrastructure or cloud. They can then customise it to their needs. Documentation for the model is linked from the release. But the notes give no parameter count, licence or benchmark figures.
The rest of the release is a patch round. It fixes a bug that had broken model loading on older versions of PyTorch. Four other targeted fixes follow. One addresses an image-assignment error in HunYuan-VL during batched generation. Another fixes a DeepGEMM issue affecting the Mega mixture-of-experts model under expert parallelism. A third resolves a from_pretrained failure on PyTorch 2.7, which now requires torch 2.8 or later for distributed checkpointing. A fourth makes DtensorShardOperation work on PyTorch 2.6 and 2.7. A separate small fix addresses a bug in auto mappings. The full changelog spans v5.20.0 to v5.20.1, with contributions from five people including Ilyas Moutawwakil and Yoni Gozlan.
Why it matters
No parameter count, licence or benchmark numbers were given. Engineers evaluating Beam for self-hosted coding agents will need to check Reflection AI's own documentation before committing infrastructure. Anyone running older PyTorch builds should take the patch fixes seriously before their next deploy.
Primary source
Transformers releases — github.comWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.