Meta Returns to Open Source With Muse Glimmer, a 30B Model Built to Run on One GPU

Meta on Monday released Muse Glimmer, a 30-billion-parameter open-weight model designed to run autonomous AI agents directly on consumer hardware, in its first fully open release since it retired the Llama family in April. The weights, published on Hugging Face under the permissive Apache 2.0 license, push agentic workloads that normally depend on cloud data centers onto high-end Macs and PCs with a single GPU.

Built for the Laptop, Not the Data Center

Muse Glimmer is a distilled version of Meta's frontier model Muse Spark 1.2, with reasoning capabilities transferred from the larger model and a lightweight companion "drafter" added to speed up generation. Quantization compresses the weights to roughly 17 gigabytes, down from more than 55 GB in full precision, so it runs on systems with 24 or 32 GB of video memory. Meta says the model handles agentic tasks such as schedule management, file organization and coding locally, with or without an internet connection, and has published custom agent scaffolds for developers.

CEO Mark Zuckerberg paired the release with a 6,500-word essay arguing that "rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it." He warned that US open-source developers face regulatory disadvantages on training-data use and distillation relative to Chinese competitors, and said Meta would also open the weights of Muse Spark 1.2 "in coming weeks" — a step that would put a US frontier-class model into open circulation, something no American lab has yet done at that tier.

An Ecosystem Push

Support is rolling out this week through Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI and OpenRouter, with llama.cpp, MLX and ExecuTorch integrations following, according to VentureBeat. Meta says it is working with AMD, Arm, Dell, Intel and Nvidia to optimize the model across devices. Chief AI Officer Alexandr Wang said Glimmer can operate as a "fully capable agent" with failure recovery on local hardware.

The release is the latest escalation in Meta's price war with the frontier labs: API pricing for its previous model, Muse Spark 1.1, came in at roughly 25 percent of what Anthropic and OpenAI charge for comparable models, Zuckerberg said last month. Meta has said nothing about training data or training code — as with most "open source" releases, only the weights are open, a nuance analysts flagged at launch, with CNBC and Bloomberg reporting the move is aimed at reassuring investors Meta remains a contender against OpenAI and Anthropic.