Meta released Muse Glimmer, a new multimodal model designed for local agentic use cases. Distilled from Muse to 30B parameters and released under the Apache 2.0 license, it's optimized for local deployment for privacy, cost reduction, or experimentation. Intended for privacy-aware applications including coding, document analysis, personal assistants, and Claw- or Hermes-like setups.
Day-0 support is included in transformers, llama.cpp, vLLM, Inference Endpoints, and other libraries. The model is available on the Hugging Face Hub.
Benchmarks show Muse Glimmer competing with or outperforming Gemma4-31B Thinking Mode and Qwen3.6-27B Thinking Mode across agentic, multimodal, and general reasoning categories. The architecture consists of a dense 30B parameter model with a 2B ViT-style vision encoder (Perception Encoder), a 28B parameter text decoder, and a speculative decoding drafter built on DFlash.