Meta's Superintelligence Labs released Muse Glimmer on Monday, and the pitch is the size: 30 billion parameters, open weights, Apache 2.0, small enough to sit on a Mac or a PC with one consumer GPU. It folds multi-step reasoning, tool calls, multimodal input, and failure recovery into a single model built for always-on local agents — the kind of thing you'd otherwise pay a cloud API to do. Weights are on Hugging Face today, with the developer docs to match. The benchmark story is aimed at people who actually ship agents: Meta reports solid full-task success on DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench, the suites that measure whether a model can work inside a scaffold, write and debug code, and finish a multi-turn request instead of stalling halfway. Optimized runtimes for llama.cpp, MLX, and ExecuTorch are promised in the days ahead, so download-to-working-agent lands in minutes. For anyone who's been keeping a local coding agent on a leash because the good models lived behind a network call, this is worth an afternoon.