source: Hugging Face Blog: Meta is back with Muse Glimmer: local, agentic, multimodal, and open source
level: technical
Meta released Muse Glimmer, a 30-billion-parameter multimodal model designed for local agentic use cases. It handles text, images, and video, and is intended for privacy-sensitive applications such as coding assistants, document analysis, and personal agents. The model is released under the Apache 2.0 license and is available on Hugging Face with immediate support in transformers, llama.cpp, vLLM, and Inference Endpoints.
Benchmarks show Muse Glimmer outperforms Gemma4-31B and Qwen3.6-27B on several agentic tasks. It scores 75.5 on MCP Atlas and 74.6 on DeepSearch QA, compared to 54.2 and 61.7 for Gemma4. On SWE-Bench Verified, it reaches 76.0, slightly behind Qwen3.6’s 77.2. The model uses a 2B-parameter vision encoder and a 28B text decoder with hybrid sliding-window and full attention, plus an optional speculative decoding drafter for faster generation.
Muse Glimmer builds on Meta’s earlier Perception Encoder architecture for vision. The text decoder alternates sliding-window and full-attention layers, using RoPE and NoPE respectively, and employs grouped-query attention to reduce memory. The vision encoder processes images and videos frame-by-frame, supporting up to 96 frames. The model is positioned for local deployment to reduce costs and improve privacy, with community-provided quantized versions already available.
why it matters: Muse Glimmer enables powerful multimodal agentic AI to run locally on consumer hardware, reducing reliance on cloud APIs and improving data privacy for developers building coding assistants, document analysis tools, and personal agents.
source: Hugging Face Blog: Meta is back with Muse Glimmer: local, agentic, multimodal, and open source