The 30-billion-parameter model is free to download under an Apache 2.0 license, and Meta says it can handle agentic tasks like scheduling and coding on a laptop or gaming PC.
The takeaway
Meta released Muse Glimmer on August 10, a 30-billion-parameter open-weight AI model licensed under Apache 2.0 that runs on a single consumer GPU. Designed for local, always-on agent tasks like scheduling, coding, and file organization, it reportedly outperforms Gemma4-31B and Qwen3.6-27B on several benchmarks, according to Meta's research blog.
Meta released Muse Glimmer on August 10, an open-weight AI model built to run entirely on a single consumer graphics card rather than in a data center. The 30-billion-parameter model is licensed under Apache 2.0, meaning developers can download, modify, and redistribute it for free, including for commercial products.
What Muse Glimmer is designed to do
Meta is positioning Glimmer as an 'agentic' model built for tasks a personal assistant might handle: drafting messages, organizing files, managing a calendar, and learning a user's habits over time. According to Meta's research blog, the model supports multi-step reasoning, function calling, error recovery when a task fails partway through, and multimodal input that combines text and images.
- Quantized to roughly 4-bit precision, keeping the model under 20GB so it fits in 24GB or 32GB of GPU memory
- Runs on both Mac and PC hardware, including gaming-grade GPUs
- Supports over 100 languages and integrates with third-party agent frameworks such as OpenClaw
- Meta says speculative decoding delivers up to a 3.1x speed increase on an Nvidia RTX 5090 and smaller gains on Apple's M-series chips
How it stacks up, according to Meta
Meta says Glimmer outperforms Google's Gemma4-31B and Alibaba's Qwen3.6-27B on several benchmarks, including DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench. Those figures come from Meta's own testing and have not yet been independently verified by outside researchers, which is standard for a same-day model launch but worth keeping in mind before treating the comparisons as settled.
Why open-weight, not open-source
Releasing model weights lets anyone run and fine-tune Glimmer locally, but Meta has not published the training data or full code used to build it — a distinction that matters for developers deciding how much they can inspect or reproduce.
The bigger picture
The release comes as Meta pushes further into what CEO Mark Zuckerberg has called personal superintelligence — AI that runs close to the user rather than purely in the cloud. Free, locally-run models also give Meta a way to compete with closed, subscription-based AI products from rivals without charging for access.
Why you can trust this article
Written and edited in-house by the THRYV Tech Desk. We do not republish or reword agency copy, and we do not invent quotes, statistics, testimonials or ratings. Where figures move frequently, we point you to the primary release rather than printing a number that will be out of date. Advertising and affiliate partnerships have no influence on our reporting — see our editorial standards, fact-checking policy and affiliate disclosure. Spotted an error? Write to newsroom@thryv-news.com.
General information only. Not personalised financial, medical or legal advice.
