Meta Releases Muse Glimmer 30B: Open Agentic Model Running Locally on Consumer GPU

Meta released a 30B open agentic model that runs on a single consumer GPU (24GB VRAM). Uses local/global attention, DFlash speculative decoding, quantization to fit 24GB. Challenges cloud-first assumption for AI agents. Privacy-sensitive workflows and offline agents become viable. This is a major signal in the open-weight model race, complementing Qwen's momentum.