Meta’s release of Muse Glimmer has brought the debate over open-weight AI back into focus, with the company making a 30-billion-parameter model available to developers that can run on a laptop or a single consumer GPU. Rishi Raj explains what the launch means
l What exactly has Meta launched?
LAST WEEK, META released Muse Glimmer, a 30-billion-parameter AI model designed specifically for agentic tasks that can run locally on consumer hardware. Unlike AI models that depend on cloud servers to process requests, Glimmer can operate on a Mac or PC within a 24-32 GB memory envelope. The company has released the model weights under an Apache 2.0 licence, allowing developers to download, run, modify and build applications around it.
l Significance of the model running on a laptop
MOST ADVANCED AI models are run in large data centres because of the computing power and memory they require. Meta says a 30-billion-parameter model running at full precision would require more than
55 GB of memory. It has used quantisation to compress Glimmer’s weights to around 4-bit precision, bringing the model below 20 GB and allowing it to run within a 24 GB or 32 GB memory envelope. Meta has tested the model on systems including Apple’s M4 Max and M5 Max and Nvidia’s RTX 5090. This means developers can potentially run the model locally without sending their data to an external cloud server or requiring a constant Internet connection. Glimmer can make a plan, call tools available to it, examine the results and continue working towards a goal.
l What does open-weight mean?
OPEN-WEIGHT MEANS Meta is releasing the trained parameters, or weights, of the AI model so developers can download and run the model themselves. The weights essentially contain what the model has learned during training. This is different from accessing an AI model through an API, where the company retains control of the model and users pay to send requests to its servers. With Glimmer, developers can take the model, run it on their own infrastructure and customise it.
l Is open-weight the same as open-source?
OPEN-SOURCE IS A broader concept that involves making the underlying source code available for inspection, modification and redistribution under an open-source licence. An AI model can be open-weight without every element of its development being open. The company may release the weights and some of the software required to run the model without releasing all of its training data, training infrastructure or complete training process. In Glimmer’s case, Meta is releasing the weights under the permissive Apache 2.0 licence, while also providing inference code and developer documentation.
Open-weight models can be downloaded and run by companies themselves, giving them greater control over data and potentially reducing dependence on cloud-based AI services.It also reflect the growing competition between US and Chinese AI developers. Chinese firms have released several powerful open-weight models, while leading US developers such as OpenAI and Anthropic have largely kept their most capable models closed.
l Zuckerberg’s views about US policy
ZUCKERBERG IS SAYING that US companies face greater restrictions on areas such as training data and AI distillation than some foreign competitors. He wants the US to reduce what he calls this additional friction if American companies are to remain competitive in open-weight AI.
l What is AI distillation & why is it controversial?
DISTILLATION IS A technique in which a smaller model learns from the outputs of a much larger and more capable model. The aim is to transfer some of the larger model’s capabilities into a smaller system that requires less computing power to operate. Meta says Glimmer was trained using outputs from its much larger Muse Spark model, followed by further training and reinforcement learning.
l Is Meta moving away from bigger AI models?
NOT NECESSARILY. THE company is doing both. Glimmer is designed to bring agentic AI to local devices, but Meta has also said it will release the weights of Muse Spark 1.2, its more advanced foundation model, in the coming weeks.
l What is the larger significance of Glimmer?
THE LAUNCH REPRESENTS a push to make AI less dependent on the cloud. If increasingly capable models can run locally, developers and businesses can potentially build applications where data remains on the user’s device, while avoiding some cloud inference costs and connectivity requirements.