Hugging Face has released Transformers v5.15.0, adding support for Meta’s Muse Glimmer multimodal model alongside a substantial sweep of attention, vision, generation, cache and quantisation changes. The headline for developers is broader local and multimodal model support. The footnote, wearing steel-capped boots, is a set of breaking changes that may affect existing code.
Hugging Face Watch analysis
What happened
The 10 August release adds Muse Glimmer, a 30B-parameter model with a vision encoder and text decoder, released under the Apache 2.0 licence. Hugging Face also added support for several other model families, including Granite variants, SKT’s A.X-K1 and A.X-K2, and Cosmos3 Edge.
For existing users, kernels are now opt-in for several linear-attention models. The cache-cropping API now expects negative relative offsets, while T5-family models can use SDPA and other attention backends. Downstream code importing certain private multimodal processor helpers will also need updating.
Our top picks
- Muse Glimmer support
Transformers can now handle Meta’s multimodal 30B model, aimed at agentic and local-use cases. - Kernel choice changes
Linear-attention models no longer select kernels automatically, so performance-sensitive setups may need explicit configuration. - Cache API adjustment
Direct cache-cropping calls must move from absolute sizes to negative relative offsets. - T5 attention backends
T5 and related families gain SDPA and other backends, potentially changing the default attention path. - Vision preprocessing
Several vision-language pipelines receive performance and correctness fixes, including image preprocessing and embedding alignment. - Generation fixes
Batched audio generation arrives for Qwen2.5/3-Omni, with further fixes for speculative decoding and model-specific crashes.
Why it matters
This is useful infrastructure work with one important caveat: upgrading is not entirely ceremonial. Teams relying on implicit kernel selection, direct cache APIs or eager-only T5 attention should inspect their configurations before treating the version bump as routine.
The upside is a wider, better-maintained runway for multimodal and local model workloads. The downside is that some of the runway markings have moved overnight.
Our read
Worth upgrading for new model support and the accumulated fixes, but stage it. Run regression tests around cache handling, T5 attention and linear-attention performance before rolling it into a shared environment.
What to watch
- Whether kernels become a required dependency for the PyTorch extras in a future release.
- Performance changes in linear-attention models after explicit kernel configuration.
- Compatibility issues in downstream packages using removed private processor helpers.
- Follow-up fixes for the new attention and cache paths.
Discussion spark: Which change in Transformers 5.15 is most likely to affect your current stack: kernel selection, cache cropping or the T5 attention backend?
Sources and evidence
- Release Release: v5.15.0 · huggingface/transformers · GitHub (2 September 2026, 14:03 UTC)
not affiliated with or endorsed by Hugging Face