Everything you need to know about Muse Glimmer and Meta’s AI plans

After appearing to abandon open-source AI earlier this year, Meta has now released a new open-source model called Muse Glimmer. Here’s everything you need to know about this latest AI model…

What on earth is Muse Glimmer? It sounds like a mid-90s indie band.

Muse Glimmer is a brand-new, 30-billion-parameter open-weight artificial intelligence model developed by Meta Superintelligence Labs. Released alongside a massive 6,000-word manifesto by Meta CEO Mark Zuckerberg, it marks Meta’s return to releasing open weights after more than a year. Distilled from Meta’s larger, proprietary Muse Spark model, Glimmer is specifically engineered not as a basic cloud chatbot, but as a local system designed to run directly on consumer hardware including personal Macs and PCs with a single graphics processing unit.

Why is Meta releasing an open-weight model right now?

The release is a direct challenge to the closed-ecosystem strategies favoured by rival labs such as OpenAI and Anthropic, which argue that cutting-edge AI models are too dangerous to be made public. Mark Zuckerberg used the launch of Muse Glimmer to champion a philosophy of “superintelligence for all,” arguing that technological power should be widely distributed rather than concentrated in the hands of a small corporate elite or government entities. By offering an advanced model that anyone can download, Meta is attempting to build a broad coalition of developers and establish a powerful “balance of power” against its competitors.

What makes Muse Glimmer different from regular chatbots?

Unlike standard language models that simply respond to a single prompt, Muse Glimmer is purpose-built for autonomous “agentic” workflows. This means it is optimised to handle complex, multi-step tasks, including writing and debugging code, executing sequential tool calls, managing memory, and automatically recovering when a process fails. It also features a dedicated perception encoder, giving it native multimodal capabilities that enable it to seamlessly interpret text, documents, charts and screenshots.

Do you need a massive data centre to run it?

No, that is the whole point of the hardware optimization behind Glimmer. Because it has been distilled down to 30 billion parameters and compressed using quantisation techniques, it can run locally on standard developer setups, such as an 18GB RAM/VRAM setup, local Mac hardware, or PCs equipped with a single consumer GPU. It also utilises techniques such as speculative decoding via a lightweight companion “drafter” model, allowing it to generate text and code significantly faster without sacrificing output quality.

What kind of licence does it use?

Breaking from some of Meta’s previous custom licensing frameworks used for earlier iterations, Muse Glimmer is released under a permissive Apache 2.0 license. This open license places virtually no restrictions on commercial use, modification or redistribution. It allows developers, indie creators and enterprises to freely build derivative software, embed it into applications and deploy it using standard open-source tools and scaffolds such as llama.cpp, Unsloth, and LM Studio.

Are there any downsides?

Yes, releasing powerful open-weight models carries significant risks that critics and safety researchers have pointed out. Because the weights are fully accessible and modifiable, bad actors can strip away guardrails, fine-tune the models for malicious purposes, such as generating malware, conducting automated cyberattacks, or even creating bioweapons, and deploy them completely unmonitored without content filters. Furthermore, running a 30-billion-parameter model locally still demands substantial hardware resources, meaning everyday users without high-end consumer GPUs or sufficient VRAM will struggle to run it efficiently on personal devices.

What else did Mark Zuckerberg talk about in his manifesto?

Alongside the model rollout, Zuckerberg published a sweeping 6,000-word essay titled “The Future is for Everyone,” which touched on everything from data centres and cybersecurity to the future of labour. He warned that heavy government regulation and policies that slow down American model releases, even by a matter of weeks,would allow foreign competitors – most notably China – to pull ahead in the global AI race. He also defended the massive infrastructure buildout of resource-hungry data centres, pointing to local community investments, and argued that while AI will inevitably cause labour market disruptions, personal AI agents will ultimately help humans adapt to new economic opportunities.