Meta has released Muse Glimmer, a 30-billion-parameter AI model that developers can download, modify and run on powerful personal computers. The move marks a renewed commitment to openly available model weights after Meta shifted its newer Muse family toward proprietary releases earlier this year.
Glimmer is available under the Apache 2.0 license, which permits commercial use, redistribution and modification. That is a more permissive arrangement than Meta’s earlier Llama license, which included usage restrictions. Carl Franzen reports for VentureBeat that the model’s weights are available through Hugging Face, with support planned across tools including Ollama, LM Studio and llama.cpp.
Meta positions Glimmer less as a conventional chatbot and more as a model for AI agents. Such systems can plan tasks, use software tools, inspect results and make corrections. The model can process text and images, allowing an agent to work with screenshots, documents and interfaces as well as written instructions.
Built for local agent workflows
Meta says Glimmer can run with 24GB of graphics memory in a quantized version. Quantization reduces the precision used to store a model, lowering memory requirements while attempting to preserve performance. This puts the model within reach of high-end consumer hardware, including PCs with Nvidia RTX 3090 or 4090 graphics cards and Macs with sufficient unified memory.
The full-precision model still requires far more memory, according to Meta. The company has also released smaller 4-bit versions and a speculative-decoding component designed to improve response speed. In speculative decoding, a smaller model proposes text that the main model checks, reducing the time needed to produce output.
Meta’s own benchmark results show Glimmer performing strongly on several tests for tool use and software development. However, competitors such as Alibaba’s Qwen3.6-27B lead on other tasks, including some computer-use and multimodal evaluations. The comparisons have not yet been independently verified.
For businesses, local deployment could offer practical benefits. Sensitive files, internal systems and work contexts can remain on an organisation’s own equipment rather than being sent continuously to a cloud AI service. It can also reduce per-token API charges. But organisations still need to pay for suitable hardware, maintenance, electricity and security controls.
Meta also warns that a locally run agent needs safeguards. Keeping data on-device does not prevent prompt injection, overly broad software permissions or unintended actions. The company recommends human approval for irreversible steps.
Zuckerberg frames open models as a political choice
The release accompanies a lengthy policy statement from chief executive Mark Zuckerberg. He argues that advanced AI should be distributed widely rather than controlled by a small number of companies or governments. In Meta’s official post, Zuckerberg describes this approach as “personal superintelligence” and says individuals should have agents that support work, learning, health, finances and creative projects.
He also defends the use of distillation, a technique in which a smaller model learns from the outputs of a more capable one. The issue has become politically sensitive as Chinese AI companies release increasingly capable open-weight models at lower costs. Some US labs have accused rivals of using their outputs to train competing systems.
Meta says it will soon release the weights for Muse Spark 1.2, a more capable model used in its Muse Code product. It has not said whether it will open the weights of its future flagship models. That decision will determine whether Glimmer represents a one-off release or a broader change in Meta’s AI strategy.
Sources
- With his long essay, Zuckerberg returns to his “open” AI arguments at an opportune time, as Chinese open models narrow the performance gap at much lower cost – Spyglass
- Meta returns to open source with Muse Glimmer, an Apache 2.0 licensed 30B parameter AI model optimized for agents – VentureBeat
- The Path to a Positive AI Future – Meta
- Mark Zuckerberg criticizes “closed” AI model makers, says AI labs are espousing a “discourse … so filled with doom”, and defends distillation as a principle – Financial Times
Stay up to date
AI for content creation: the latest tools, tips and trends. Every two weeks in your inbox:
