Meta releases Muse Glimmer, a 30B open model for local agents

Meta just released an open 30-billion parameter model built to run continuous AI agents on a single desktop GPU.

Meta Superintelligence Labs open-sourced Muse Glimmer under an Apache 2.0 license. The 30B multimodal model focuses on local, always-on tasks like code generation, document processing, and function calling. Using 4-bit quantization, the language model shrinks under 20 GB to run inside a 24 GB or 32 GB RAM envelope. It also uses a small companion model called DFlash for speculative decoding, which speeds up long reasoning chains.

Why it matters: Autonomous agents usually require cloud servers and constant internet access to run long tasks. Glimmer shifts that workload offline, giving you local tool execution, vision, and error-recovery without sending personal context to an API. Meta's benchmarks show strong agentic performance compared to size rivals like Gemma4-31B and Qwen3.6-27B.

The open weights are live on Hugging Face today. Optimized integrations for llama.cpp, MLX, and ExecuTorch arrive in the coming days.

Local AI just got a lot more practical—and your API bills might finally catch a break.