Meta is back with Muse Glimmer: local, agentic, multimodal, and open source

Why it matters

Meta has released Muse Glimmer, an open-source multimodal model that Hugging Face supported from launch day across its ecosystem, including transformers, llama.cpp, vLLM, and Inference Endpoints. The model pairs a Perception Encoder vision component, previously introduced by Meta as a backbone for spatial and multimodal tasks, with a text decoder that uses hybrid attention mixing sliding-window layers with periodic full-attention layers, plus grouped-query attention to shrink the key-value cache and speed generation. An optional speculative decoding drafter built on DFlash trades extra memory for faster output, and Hugging Face found it especially effective for structured content such as code. The same loading code runs on NVIDIA, AMD, and Intel GPUs, with automatic device placement. The release matters because it gives the open-source community a capable, openly available multimodal model with immediate tooling support, lowering the barrier for developers building image, video, and text applications locally.