What Meta's Muse Glimmer means for you

Meta has released a new AI model that you can download, run on your own computer, and even build products with, free of charge.

On 10 August 2026, Meta published the "weights" (the trained numbers that make a model work) for a model called Muse Glimmer on Hugging Face, a site where AI models are shared. It carries the Apache 2.0 licence, which means anyone can use it commercially, change it, or redistribute it.

Muse Glimmer has 30 billion parameters, the internal settings that determine how a model behaves, making it a mid-sized model by today's standards. It was distilled, or compressed and simplified, from Muse Spark 1.2, Meta's more powerful closed model released on 5 August. This is Meta's first fully open model since it retired its well-known Llama line earlier in 2026 in favour of the proprietary Muse Spark. Meta chief executive Mark Zuckerberg also said the company will open-source Muse Spark 1.2 itself. He described Muse Glimmer as a great 30B parameter dense model that can run locally.

The full version of the model is 59.6 GB, but a compressed "4-bit" version comes in under 20 GB, small enough to run on a single 24 or 32 GB graphics card or a Mac, with less than 1% loss in accuracy. It also ships with a small helper model, called a drafter, that speeds up responses: Meta reports the model runs 3.1 times faster on an Nvidia RTX 5090 graphics card, and 1.5 to 1.8 times faster on Apple's M4 and M5 Max chips.

The model can handle both text and images, understands longer conversations (over 131,000 tokens, roughly the length of a long book), and was specifically trained to use external tools, call software functions, and keep working even if one of those tools fails. On a benchmark called MCP Atlas, which tests this kind of tool use, it scored 75.5, ahead of Google's Gemma 4 31B (54.2) and Alibaba's Qwen 3.6 27B (62.5).

If you are a developer or a hobbyist with a decent gaming PC or a recent Mac, you can now run a capable AI model entirely on your own hardware rather than relying on a cloud service. That means no per-use fees to a company, and more control over your data since nothing needs to leave your machine.

Because it is built for "agents" — AI systems that carry out multi-step tasks using outside tools — Muse Glimmer is aimed less at casual chatbot use and more at people building assistants that can, for example, search the web, call APIs or automate tasks.

The model weights are available now on Hugging Face under the Apache 2.0 licence, meaning you or a developer you rely on can download, adapt and deploy it, including for commercial products.