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

  • Optimized for end-to-end agentic task completion on benchmarks including DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench.
  • Supports reliable tool use with precise function call schemas across extended workflows.
  • Capable of multi-step reasoning over long horizons to sustain coherent plans in complex tasks.
  • Functions as a vision model, enabling image description and analysis alongside text-based interactions.

The model is designed for local deployment on machines with 32 GB or more of RAM, allowing sufficient memory for other applications while handling full-task workflows.