Meta releases Muse Glimmer for local AI agents
Meta’s 30B open-weight model is built to keep agent work running on consumer hardware without a constant cloud link.
Muse Glimmer is aimed at local agents, coding tools, function calling, and model-based evaluation on a Mac or PC with a single consumer GPU. Meta says it handles multi-step reasoning, tool failures, interleaved text and images, and more than 100 languages. Quantized builds can bring the language model under 20 GB, with reported decode-speed gains from a companion drafter. Weights and documentation are on HuggingFace now, with llama.cpp, MLX, ExecuTorch, Ollama, LM Studio, and other support planned. TestingCatalog's note
Muse Glimmer is aimed at local agents, coding tools, function calling, and model-based evaluation on a Mac or PC with a single consumer GPU. Meta says it handles multi-step reasoning, tool failures, interleaved text and images, and more than 100 languages. Quantized builds can bring the language model under 20 GB, with reported decode-speed gains from a companion drafter. Weights and documentation are on HuggingFace now, with llama.cpp, MLX, ExecuTorch, Ollama, LM Studio, and other support planned. TestingCatalog's note
score 7