Meta releases Muse Glimmer, 30B open-source AI model

Meta just flipped the switch on Muse Glimmer, a 30-billion-parameter AI model that runs autonomous agents on a local PC or Mac — no cloud required. The company pushed the open-weight weights to Hugging Face today under the Apache 2.0 license, marking its first fully open release since it moved its flagship Llama models behind a proprietary license last spring.

After years of criticism for restrictive terms like the 700-million-user cap in Llama’s community license, Meta is now shipping under Apache 2.0, which lets anyone use, modify and sell the model without strings. The timing is deliberate: after shifting Llama to closed-source Muse Spark, Meta is quietly reopening with a more permissive option, signaling a strategic pivot back toward open development while still keeping its biggest commercial models locked down.

The 30B parameter model is tuned for “agentic” tasks — planning, tool use, multi-step workflows — that normally eat cloud budgets. Meta says it can run on high-end consumer GPUs, pushing inference workloads from data centers back onto desktops. Early partners like Ollama, LM Studio, vLLM and Together AI are rolling out immediate support, letting users spin up agents with a few clicks. A company spokesperson said the release reflects “a renewed commitment to open innovation while protecting our most valuable models.”

What matters next is whether developers treat Apache 2.0 as a green light to build commercial products around Glimmer. If adoption spreads, Meta could ease pressure on cloud costs and broaden AI access — but it also keeps its crown-jewel models proprietary, drawing a clear line between open tools and high-margin products.