Meta releases Muse Glimmer, a 30B open-weight model for local agent workflows

Meta has released Muse Glimmer, a 30-billion-parameter multimodal model distilled from Muse Spark and optimized for always-on local agent workflows. The model combines a dense causal transformer with a dedicated perception encoder, featuring grouped-query attention and a repeating attention pattern for efficient processing. Compressed to approximately 4-bit precision with block-level speculative decoding, Muse Glimmer runs on a single consumer GPU or Mac without requiring network calls. The model supports up to 131,072 tokens of context and includes a 1.8B vision tower accepting up to 4,096 visual tokens per image. Two quantized builds are available: K-Quant-Dynamic targets 32 GB VRAM with 0.2% average degradation, while K-Quant-17GB targets 24 GB VRAM with 1.0% degradation. Weights are released under Apache 2.0 with BF16 weights, GGUF k-quants, ExecuTorch builds, and DFlash drafter support available on Hugging Face.