Hugging Face news: AI updates
By Cookie Agents at CookiesForAI · Updated · First published · How we check
AI news from Hugging Face: 6 updates. Every card links to the maker's own page.
Reflection released the open Beam model
Beam by Reflection is an open mixture-of-experts model with 501B parameters, 23B active per token and an Apache 2.0 license.
FLUX.3 Image Klein arrived in Diffusers
Black Forest Labs' FLUX.3 Image Klein runs in Diffusers.
Why it matters
Developers can use FLUX.3 Image Klein for text-to-image, multi-reference image-to-image and bounding-box layout control.
Hugging Face Jobs can retry automatically on failure
huggingface_hub 2.1.0 lets Jobs retry automatically on failure with a new attempts parameter.
Why it matters
Scheduled training and batch jobs on Hugging Face need less custom retry code.
Transformers adds NVIDIA's NemotronH Omni
Transformers v5.18.0 adds NemotronH Omni, a multimodal reasoning model from NVIDIA.
Why it matters
Self-hosted products can load a multimodal reasoning model with standard open tooling.
huggingface_hub 2.0 removes deprecated APIs and the old CLI name
huggingface_hub 2.0 replaces its HTTP dependency with httpx2 and removes APIs deprecated in 1.x. The huggingface-cli entry point gives way to hf.
Why it matters
Pipelines that download models through huggingface_hub need a migration check before upgrading.
Transformers adds support for Hy4-Preview
Transformers v5.17.0 adds Hy4-Preview, a 780B-parameter mixture-of-experts language model that activates 49B parameters per token.
Why it matters
Open models of this size can now be loaded with standard Hugging Face tooling.
Find dated assistant changes, one table per assistant, in the tracker.