Meta is back with Muse Glimmer: local, agentic, multimodal, and open source
huggingface ·
Meta has introduced Muse Glimmer, a 30-billion parameter multimodal model specifically engineered for local and agentic workflows. Distilled from the original Muse architecture and distributed under an Apache 2.0 license, the system pairs a 2-billion parameter vision encoder with a 28-billion parameter text decoder. Its technical design incorporates hybrid attention mechanisms, grouped-query attention to reduce memory overhead, and an optional DFlash speculative decoding module that accelerates structured outputs like code. The model natively handles both static images and video clips, processing up to 96 evenly sampled frames per sequence without relying on audio cues. Independent benchmarking reveals competitive performance against similarly sized competitors across agentic planning, software engineering, multimodal reasoning, and safety assessments. Developers can deploy the model immediately using day-zero integration in major inference libraries such as Hugging Face Transformers, llama.cpp, and vLLM. By emphasizing privacy preservation and reduced operational costs, Muse Glimmer targets practical applications including automated coding assistants, document analysis, and localized personal agents.
huggingface