Meta has released Muse Glimmer, a new 30-billion parameter multimodal model available under an Apache 2.0 license. The release marks Meta's return to the open weights space with a focus on local deployment and agentic capabilities.

  • Optimized for end-to-end agentic task completion, achieving strong success rates on benchmarks including DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench.
  • Designed for reliable tool use, handling wide ranges of function calls with precise schemas throughout extended workflows.
  • Capable of multi-step reasoning, sustaining coherent plans across complex, long-horizon tasks.
  • Functions as a vision model, capable of processing and describing images alongside text inputs.

The model is positioned to support local inference on machines with 32 GB or more of RAM, allowing users to run other applications simultaneously while utilizing its agentic features.