Meta releases Muse Glimmer, a 30B open source multimodal model
Drafted by our automated systems from the sources listed at the end of this article, and reviewed before publication. Every fact here should be traceable to one of those links; we publish no benchmark, price or figure that a linked source does not state. Where a company makes a claim about its own product, we report it as their claim.
Meta has released a new model called Muse Glimmer, described as local, agentic, multimodal, and open source. According to Hugging Face, day-0 support for the model is available across transformers, llama.cpp, vLLM, Inference Endpoints, and other libraries. The model is available on the Hugging Face Hub.
Model architecture and optional speculative decoding
Muse Glimmer is a dense 30B parameter model combining a 2B ViT-style vision encoder known as the Perception Encoder with a 28B parameter text decoder. The vision encoder handles images and videos by patchifying frames and processing them through 50 layers. The text decoder features hybrid attention, gated grouped-query attention, and Q-K normalization. Additionally, Muse Glimmer includes a speculative decoding drafter implemented on DFlash. Usage of this module is optional and provides faster generation for structured content like coding in exchange for some memory cost.
Is it worth it?
Muse Glimmer is worth trying for developers looking for an open source, local multimodal model with immediate library support. Readers should verify how the hardware memory requirements fit their local setups, particularly when enabling the optional DFlash speculative decoding drafter.
What this changes for tools we list
- Hugging Face — Muse Glimmer is now available on the Hugging Face Hub with day-0 support across transformers and Inference Endpoints.
Sources
- Meta is back with Muse Glimmer: local, agentic, multimodal, and open source — Hugging Face, 10 August 2026