Hugging Face v5.15.0 adds Meta's 30B Muse Glimmer for agents

Hugging Face shipped Transformers v5.15.0, headlined by support for Meta's brand-new Muse Glimmer model. This dense 30B-parameter multimodal model combines a 2B ViT-based Perception Encoder for vision with a 28B text decoder, and it's released under the permissive Apache 2.0 license. Designed specifically for agentic use cases, Muse Glimmer can run locally for privacy-sensitive applications like coding assistants, document analysis, and personal AI setups similar to Claw or Hermes architectures. It's a significant addition because it offers a fully open-weight, agent-capable model that developers can deploy without relying on cloud APIs.

Beyond Muse Glimmer, this release adds support for several other architectures: IBM's GraniteSWA and GraniteMoeSWA models with sliding window attention, SKT's A.X-K1 and A.X-K2, and Cosmos3 Edge. But v5.15.0 also introduces breaking changes developers must note. Kernels for linear-attention models (Mamba, GDN, Conv-only) are now opt-in instead of automatically selected. The cache cropping API now requires negative relative offsets instead of absolute sizes. T5 and its family (MT5, LongT5) now support SDPA and other attention backends, with eager attention no longer the default. Finally, several private helper functions were removed from multimodal processors. The release also includes extensive attention fixes, including MLA cache compression and Flash Attention optimizations for vision models, along with improved MPS support.