Meta released Muse Glimmer on Monday, an open-weight model small enough to run on one GPU. The 30-billion-parameter version is on Hugging Face now, with optimised builds landing in llama.cpp so the trip from download to working agent takes minutes rather than an afternoon.
It is distilled from Muse Spark 1.2, Meta’s closed flagship, and aimed squarely at agent work: scheduling, file management, tool use, multi-step reasoning and recovering when a step fails. It takes multimodal input and was trained across more than 100 languages.
The positioning matters more than the benchmark scores. Almost all agent work today runs in rented data centres, metered per token. A capable model that runs locally changes that arithmetic for anyone leaving agents running continuously rather than poking at them interactively.
Mark Zuckerberg put the case plainly: rather than centralising superintelligence, he wrote, we should distribute it widely. Worth noting the asterisk, though. Spark still trails OpenAI and Anthropic at the top of the range, and “open source” is doing some heavy lifting here, since what Meta actually released is the weights rather than the training data.
