Meta's Muse Glimmer: A 30B Open-Source Model That Runs on Your Device

Meta Muse Glimmer – open weights 30B local coding model

Meta Superintelligence Labs unveils Muse Glimmer, a 30-billion-parameter model optimized for local agent workflows. Open-sourced under Apache 2.0, it runs on a single consumer GPU, enabling on-device agents, function calling, coding, and LLM-as-a-judge. The model is trained via distillation from a larger teacher and features quantization and speculative decoding for speed. It excels in agentic benchmarks, tool use, and multimodal reasoning, with integrations for llama.cpp, MLX, and ExecuTorch.