Its research page says the model combines multi-step reasoning, tool use, and failure recovery, and runs locally without cloud infrastructure. It accepts text and image inputs and produces text outputs; Meta says it is quantized to approximately 4-bit precision for single-GPU inference and released under Apache 2.0. Meta also reports state-of-the-art results for its size class on agentic benchmarks, including tool use, multi-step planning, and task completion. The announcement does not provide benchmark conditions or independent validation in the supplied material.
Meta released Muse Glimmer 30B , an open-weight agentic model under Apache 2.0.
In one minute
Meta introduced Muse Glimmer 30B, an open-weight causal language model with a perception encoder designed for autonomous agentic tasks on consumer hardware.