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

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.

Takeaways

1. 1

   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.

2. 2

   Its 2B vision encoder handles images and videos; video processing targets 2 frames per second and caps clips at 96 evenly sampled frames.

3. 3

   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.