Transformers v5.15.0 Release: Meta's Muse Glimmer Multimodal Model and Major Performance Boosts
Overview
Hugging Face released Transformers v5.15.0, a major update adding support for new models and introducing significant performance and infrastructure improvements. The highlight is Meta's Muse Glimmer, a multimodal model designed for agentic applications, along with support for GraniteSWA, A.X-K1/K2, and Cosmos3 Edge.
FAQ
What is Muse Glimmer?
Muse Glimmer is Meta's 30B-parameter multimodal model (2B vision encoder + 28B text decoder), designed for agentic applications like coding, document analysis, and personal assistants, licensed under Apache 2.0.
What are the breaking changes in v5.15.0?
Key breaking changes: kernels are now opt-in for linear attention models, cache cropping API now only accepts negative values, T5 supports SDPA, and some private helper functions were removed from multimodal processors.
Can Muse Glimmer be deployed locally?
Yes, thanks to its Apache 2.0 license and 30B size, it can be deployed in local setups for privacy-sensitive applications, making it suitable for enterprises in the MENA region.
What are the main improvements for developers in the region?
Performance improvements like faster image preprocessing, expanded FSDP support, security fixes, and broader hardware support (XPU, MPS), facilitating model development and deployment.
Source: Hugging Face Transformers Releases
AI-assisted content, human-reviewed.