Meta returns to open source with Muse Glimmer

Meta has released Muse Glimmer, a new open-weights model with 30 billion parameters under the permissive Apache 2.0 license. After stepping back from open Llama releases, observers read the move as Meta’s return to an open-source strategy – this time with a clear focus: local, agentic applications.

Technically, the model offers a 131,072-token context window, handles text and image input, and supports more than 100 languages. In its 4-bit quantized variant it runs in roughly 24 gigabytes of VRAM on a single consumer GPU – with only about one percent accuracy loss, according to Meta. With speculative decoding (DFlash), it reaches up to a 3.1x speedup on an RTX 5090.

On agent benchmarks, Meta says the model leads its size class: it scores 75.5 on the MCP Atlas benchmark, well ahead of comparable models. It trails, however, on classic computer-vision tasks.

Strategically, Meta is targeting industries that can’t or won’t send AI workloads to the cloud: healthcare, law firms, financial services and other regulated sectors that need data residency and offline operation. Always-on agents on your own hardware – that is the scenario Muse Glimmer was built for.