Meta released Muse Glimmer, a 30-billion-parameter open-weight model built to run AI agents locally on a single consumer GPU or a Mac, without a network connection. The model is available now on Hugging Face under an Apache 2.0 license, according to Meta’s developer documentation.
Built for always-on agents
Glimmer is distilled from Muse Spark, the closed model Meta’s superintelligence lab debuted in April, and is tuned specifically for always-on local agents — software that keeps state across restarts, calls external tools reliably, and recovers from failures without cloud infrastructure. Meta compresses the model to roughly 4-bit precision and pairs it with a lightweight “drafter” model based on a block-diffusion technique called DFlash, which predicts 16-token blocks in a single pass to speed up generation without changing output quality.
On Meta’s published benchmarks, Glimmer scores 76.0 on SWE-Bench Verified and 51.2 on the harder SWE-Bench Pro for coding tasks, 75.5 on MCP Atlas and 65.9 on OSWorld-Verified for agentic tasks, and 94.7 on AIME 2026 for math reasoning. The model also handles images and text across more than 100 languages.
It is the third Muse-branded release since April, following the coding-focused Muse Code terminal agent and the Muse Spark 1.1 update.
Open tier, closed core
The release shows where Meta draws the line between open-weight and proprietary AI. Muse Spark, the larger model Glimmer was distilled from, stays closed. In an essay accompanying the launch, chief executive Mark Zuckerberg wrote that distributing superintelligence widely gives “every person the ability to direct it,” rather than concentrating the technology inside a handful of companies. Meta’s chief AI officer, Alexandr Wang, said open weights for a larger model, Muse Spark 1.2, are coming soon.
Glimmer follows a string of small, on-device agent models released by rival labs this month, including Liquid AI’s LFM2.5-2.6B, as vendors compete to put capable AI agents directly on consumer hardware instead of the cloud.