Meta has released Muse Glimmer, a dense 30B parameter multimodal model available as open source on the Hugging Face Hub. The architecture consists of a 28B text decoder and a 2B ViT-style Perception Encoder for vision tasks.

  • The text decoder utilizes hybrid attention with sliding windows and full attention layers, along with Gated Grouped-Query Attention to reduce KV-cache memory by 16x.
  • The model supports video inference by processing frames at 2 frames per second, capping clips at 96 frames.
  • Day-0 support is provided for transformers, llama.cpp, vLLM, and Inference Endpoints.
  • An optional DFlash speculative decoding drafter is included to accelerate generation, particularly for structured content like coding.

The release enables local deployment of agentic, multimodal capabilities with broad hardware compatibility across NVIDIA, AMD, and Intel accelerators.