Meta launches Muse Glimmer for on-device AI
Meta Platforms released a new open-weight AI model on Monday, named Muse Glimmer, designed for agentic tasks that can run locally on a Mac or PC with a single graphics card. The company made the model's weights available under an Apache 2.0 licence, allowing developers to download, modify and build applications around it.
The 30-billion-parameter model is built to operate within a 24-32 GB memory envelope. According to Meta, running the model at full precision would require more than 55 GB of memory, so the company used quantisation to compress the weights to around 4-bit precision, bringing the model below 20 GB. It has been tested on systems including Apple's M4 Max and M5 Max and Nvidia's RTX 5090.
Because the model runs locally, developers can potentially avoid sending data to external cloud servers or requiring a constant internet connection. Glimmer can make a plan, call tools, examine results and continue working toward a goal, according to the company.
Meta says Glimmer was trained using outputs from its much larger Muse Spark model, followed by further training and reinforcement learning.
Zuckerberg's case for open-weight AI
In a video post accompanying a 14-page essay titled "The Future is for Everyone," Meta CEO Mark Zuckerberg championed spreading AI rather than leaving it in the hands of a few. He said the notion that AI is so dangerous that the only safe path is extreme concentration of power "seems inherently problematic."
Zuckerberg also called for lower US barriers for open-source AI to better compete with Chinese rivals. He argued that American labs face additional restrictions on training data compared with foreign counterparts, saying: "Foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data." He said the US needed to rethink policies on data use and distillation to remain competitive.
He also said infrastructure buildout is more difficult in the US than in China, and unveiled a $1 billion fund to support communities affected by Meta's data-center build-out.
The open-weight landscape
Open-weight models are typically cheaper than leading models from the likes of OpenAI and Anthropic, and come with publicly accessible core components for easy customisation, unlike closed models that companies keep under their control. Meta's release reignites a debate that has grown sharper as businesses worry about ballooning AI costs and recent cybersecurity incidents involving models from Anthropic, OpenAI and Meta.
Dawn reports that Hugging Face, an AI coding collaboration, was hacked by a rogue OpenAI model and was forced to use a Chinese open-weight model to defend against the attack because closed-source models have curbs on cybersecurity use. Chinese startups are leading the race for open-weight models, with Moonshot's Kimi K3, Alibaba's Qwen3.8-Max and DeepSeek's V4-Flash rivaling the performance of top US systems. By contrast, the leading models from US developers OpenAI, Anthropic and Alphabet's Google are closed-weight.
Company plans and market reaction
Meta plans to release the weights of Muse Spark 1.2, its most advanced model, in the coming weeks, and the company said more large models are coming soon. Zuckerberg said "We've got even bigger models that are coming soon" in the video accompanying his essay.
Meta will also implement a governance structure giving its independent directors the power to approve safety criteria for releasing models.
Shares of Meta, which have fallen about 10 percent so far this year, were up nearly 3 percent in premarket trading on Monday.