Meta Superintelligence Labs today introduced Muse Glimmer, a 30-billion-parameter model released under a permissive Apache 2.0 license. Designed for always‑on local agent workflows, it fits on a Mac or PC with a single consumer GPU, enabling agents that manage schedules, draft messages, and run code without cloud dependence.

To keep the model lightweight, the team applied 4‑bit quantization, shrinking the weights to under 20 GB and leaving room for the 24‑32 GB working memory needed for KV cache, perception encoder, and speculative decoding. The companion DFlash drafter proposes token blocks, letting the main model verify them in parallel, which speeds generation while preserving quality.

Muse Glimmer excels on key agentic benchmarks—Deep Search QA, MCP‑Atlas, 𝛕‑Bench, and SWE‑Bench—showing strong tool use, multimodal input, and failure recovery. It supports over 100 languages, works with Open Claw orchestration, and offers controllable reasoning intensity.

The weights are available on Hugging Face, with integrations for llama.cpp, MLX, ExecuTorch, and partners like Ollama, LM Studio, and Together AI. Documentation on Meta’s AI Developer Center helps developers deploy, customize, and build responsible local agents.