---
title: "Meta is back with Muse Glimmer: local, agentic, multimodal, and open source"
description: "Meta’s Muse Glimmer is a 30B-parameter multimodal model released under the Apache 2.0 license for local, agentic applications including coding, document analysis, and personal assistants. Its dense architecture combines a 2B ViT-style perception encoder with a 28B text decoder, while an optional DFlash speculative-decoding drafter targets faster generation for structured content. The decoder uses alternating sliding-window and full-attention layers, grouped-query attention, and query-key normalization with extra query scaling. Day-zero support is available in Transformers, llama.cpp, vLLM, Inference Endpoints, and other libraries, with examples for text, images, videos, deployment, and agent workflows. Published benchmarks show varied results: Muse Glimmer leads on MCP Atlas and SWE-Bench Pro, while Qwen3.6 leads on OSWorld-Verified and SWE-Bench Verified."
---

# Meta is back with Muse Glimmer: local, agentic, multimodal, and open source

[Hugging Face](https://yomu.fyi/company/hugging-face) · Pedro Cuenca, merve, ben burtenshaw, Aritra Roy Gosthipaty · Aug 10, 2026

**Type:** Announcement

## Summary

Meta’s Muse Glimmer is a 30B-parameter multimodal model released under the Apache 2.0 license for local, agentic applications including coding, document analysis, and personal assistants. Its dense architecture combines a 2B ViT-style perception encoder with a 28B text decoder, while an optional DFlash speculative-decoding drafter targets faster generation for structured content. The decoder uses alternating sliding-window and full-attention layers, grouped-query attention, and query-key normalization with extra query scaling. Day-zero support is available in Transformers, llama.cpp, vLLM, Inference Endpoints, and other libraries, with examples for text, images, videos, deployment, and agent workflows. Published benchmarks show varied results: Muse Glimmer leads on MCP Atlas and SWE-Bench Pro, while Qwen3.6 leads on OSWorld-Verified and SWE-Bench Verified.

## Context

The model is intended for local agentic use cases, including privacy-aware coding, document analysis, personal assistants, and Claw- or Hermes-like setups. Local deployment is presented as a way to support privacy, reduce costs, or enable experimentation.

## Approach / What changed

Muse Glimmer uses a dense 30B-parameter multimodal architecture with a 2B perception encoder and 28B text decoder. Its implementation is supported across Transformers, llama.cpp, vLLM, Inference Endpoints, and other libraries, with an optional DFlash speculative-decoding drafter and documented workflows for local inference, protected endpoints, agent integration, self-optimization, and Hub research.

## Takeaways

- Muse Glimmer’s text decoder has 52 layers arranged as 13 repetitions of three 2,048-token sliding-window attention layers followed by one full-attention layer.
- Its 2B vision encoder handles images and videos; video processing targets 2 frames per second and caps clips at 96 evenly sampled frames.
- In the published comparisons, Muse Glimmer scores 75.5 on MCP Atlas and 51.2 on SWE-Bench Pro, while Qwen3.6 scores 75.6 on OSWorld-Verified and 77.2 on SWE-Bench Verified.

**Tags:** [Deployment](https://yomu.fyi/topic/deployment), [LLMs](https://yomu.fyi/topic/llm), [Performance](https://yomu.fyi/topic/performance), [Privacy](https://yomu.fyi/topic/privacy)

- Source: [Hugging Face](https://huggingface.co/blog/muse-glimmer)
- Source URL: https://huggingface.co/blog/muse-glimmer
- Ingested by Yomu: 2026-08-29T05:40:51.287Z

[Read original post](https://huggingface.co/blog/muse-glimmer)