Meta Muse Glimmer: 30B Open Model Runs Locally
Meta released Muse Glimmer on 10 August 2026: roughly 29.6B parameters, Apache 2.0, and small enough to run on a single consumer GPU. Here is what actually shipped and what it means if you build with AI.
Meta Muse Glimmer landed on 10 August 2026: a roughly 29.6B-parameter multimodal model published under Apache 2.0 and built specifically for agent work that runs on your own hardware. Meta's announcement frames it as an on-device model, and the short version is that this is the first Meta release in a while where the weights are genuinely permissive, and the quantized build fits on a single consumer GPU.
That combination is the story. Open weights are not new. Open weights with an unrestricted commercial licence, agent-grade tool calling, and a memory footprint under 20GB is a different proposition, because it moves a class of work off the API bill and onto a machine you already own.
What Meta Muse Glimmer actually is
According to the model card, Muse Glimmer is a dense causal transformer, 52 layers with a 6656 hidden dimension, paired with a ViT-G/14 perception encoder of about 1.8B parameters for vision. It was distilled from the larger Muse Spark system rather than trained from scratch. Context length is listed at 131,072 tokens or more, with sliding window attention of 2048. Training data runs through 4 January 2026, and covers more than 100 languages.
The published benchmark figures put it at 76.0% on SWE-Bench Verified, 94.7% on AIME 2026, 83.5% on GPQA Diamond, 75.5% on MCP Atlas, and 74.6% on DeepSearch QA. Meta's own comparison set is Gemma4-31B and Qwen3.6-27B, which is the right peer group and worth remembering when you see the numbers quoted without it. Benchmark scores are a starting point for a shortlist, not a verdict, and it is worth knowing what an AI benchmark measures before you weight them heavily.
The hardware tiers, in plain terms
The model card lists three memory configurations, and this is the part that decides whether the release is relevant to you:
Build | VRAM needed | Roughly what that means |
|---|---|---|
Full precision | 64GB | Workstation or rented cloud GPU, not a laptop |
K-Quant-Dynamic | 32GB | High-end desktop card or an Apple silicon machine with a large unified memory pool |
K-Quant-17GB | 24GB | A single mainstream consumer GPU |
Meta reports speculative decoding gains of 3.1 times on an RTX 5090, 1.8 times on an M5 Max, and 1.5 times on an M4 Max. Those are throughput multipliers on top of the quantized build, not accuracy claims, and they are the reason the 24GB tier is usable rather than merely possible.
Why the licence matters more than the benchmark
Apache 2.0 is the headline that will age best. It is an OSI-approved licence with no user-count ceiling, no acceptable-use policy incorporated by reference, and no clause restricting what you train on the outputs. That is a materially different position from the community licences attached to several previous open-weight releases, and it removes the legal review step that usually stalls adoption inside small companies. If you are weighing this up, the differences between open-weight and closed models are as much contractual as technical.
Meta's developer page lists the distribution formats: BF16 weights, GGUF k-quants, ExecuTorch builds, and a drafter model for the speculative decoding path.
Who this changes something for
Anyone processing data they would rather not send to a third-party API. The privacy argument for local inference stops requiring a capability sacrifice at this size.
Small teams with a spiky workload. A fixed hardware cost beats per-token billing once usage is both high and predictable.
Builders in regions or industries with data residency rules, where the model running on your own machine is the compliance answer.
Anyone who has been burned by a hosted model changing underneath them. Weights on disk do not get deprecated.
It changes less for people running light, occasional workloads. A 24GB card plus the electricity to keep it busy is worse economics than an API call if you are making a few hundred requests a day, and frontier hosted models are still ahead on the hardest reasoning work.
FAQ
Is Muse Glimmer free for commercial use?
Yes. Apache 2.0 permits commercial use, modification, and redistribution, with attribution and a notice of changes. There is no monthly-active-user threshold and no separate agreement to negotiate.
Can it run on a laptop?
On an Apple silicon laptop with 32GB or more of unified memory, yes, via the K-Quant builds. On a typical 16GB Windows laptop with integrated graphics, no.
How does it compare to frontier hosted models?
It is not intended to. Meta benchmarks it against Gemma4-31B and Qwen3.6-27B, models in the same size class. Treat it as the strongest thing you can run yourself, not a replacement for the largest hosted models.
Where do I get the weights?
The Hugging Face collection under meta-models carries BF16, GGUF, and ExecuTorch builds. If you have not run a model locally before, the practical steps are the same as running an AI coding model locally.
Releases at this size arrive every few weeks now, and most of them do not matter. This one does, mainly because of the licence. Keeping a filter for that signal is most of staying current with AI news without reading everything.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
