Meta's Muse Glimmer: open-source 30B multimodal model for local agents

Meta has unveiled Muse Glimmer, a 30B-parameter multimodal LLM designed specifically for local agentic workloads. Distilled from the larger Muse model and released under the permissive Apache 2.0 license, it targets privacy-sensitive applications like coding, document analysis, and personal assistants. The model is available on the Hugging Face Hub with day-0 integration across transformers, llama.cpp, vLLM, Inference Endpoints, and other libraries. Meta also published a drafter module built on DFlash for optional speculative decoding, which speeds up structured generation tasks such as code completion at the cost of extra memory.

Architecturally, Muse Glimmer is a dense model with a 2B ViT-style Perception Encoder for both images and videos, coupled with a 28B text decoder. The decoder uses a hybrid attention pattern alternating three sliding-window layers (2,048 tokens, RoPE) with one full-attention layer using NoPE, repeated across 52 layers. Gated Grouped-Query Attention shares each key-value head across 16 query heads, cutting KV-cache memory by 16x and reducing inference costs. In benchmarks, it leads agentic reasoning on MCP Atlas (75.5 vs. 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B) and scores 76.0 on SWE-Bench Verified, while delivering strong multimodal results like 78.8 on Charxiv Reasoning. This combination of open licensing, efficient architecture, and on-device capability makes Muse Glimmer a compelling choice for agent developers who need data privacy, lower cloud costs, or full control over their stack.