Meta
Meta releases open-weight Muse Glimmer, its first since Llama
The 30-billion-parameter model runs local AI agents on a single consumer GPU under the Apache 2.0 licence.
The answer
Meta open-sourced Muse Glimmer, a 30B agent model, on 10 August.
What happened: Meta published the weights of Muse Glimmer on Hugging Face on 10 August, under the Apache 2.0 licence. That allows commercial use, modification and redistribution.
The numbers: The model has 30 billion parameters. Full BF16 weights run to 59.6 GB; a 4-bit version comes in under 20 GB and runs on a 24 or 32 GB GPU or a Mac, with under 1% accuracy loss. Context length runs to 131,072+ tokens.
The details: Muse Glimmer was built by Meta Superintelligence Labs, distilled from Meta's closed flagship Muse Spark 1.2, which launched 5 August. It handles text and images via an 1.8-billion-parameter vision encoder and is trained for multi-step tool use, function calling and recovering when an API call fails. It ships with a small "DFlash" drafter for speculative decoding, giving reported speed-ups of 3.1x on an RTX 5090, 1.5x on an M4 Max and 1.8x on an M5 Max.
The context: This is Meta's first fully open model since it retired the Llama line in favour of the proprietary Muse Spark earlier in 2026.
Why it matters: Meta is betting that open, locally runnable models will win developers back for agent workloads that don't need a data-centre GPU.
The rivals: On the MCP Atlas benchmark, Muse Glimmer scored 75.5, ahead of Gemma 4 31B (54.2) and Qwen 3.6 27B (62.5). Its Artificial Analysis Intelligence Index score is 35.
In their words: Mark Zuckerberg said: "Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally."
What's next: Zuckerberg said Meta will also open-source Muse Spark 1.2, the closed flagship that Muse Glimmer was distilled from.
Details from gHacks.