Muse Glimmer Explained: Meta's Apache 2.0 Model and How to Run It on a Mac mini

On August 10, 2026, Meta released Muse Glimmer, a 30-billion-parameter AI model you can download and run on your own computer, under the Apache 2.0 license. That’s the same permissive license behind a huge share of the open-source software world, and according to Artificial Analysis it’s the first time Meta has used it for a model. Every earlier open Meta model shipped under one of its own Llama licenses, which came with a user cap and a carve-out for the European Union.

Glimmer is built for agent work (tasks with several steps and tool calls), and Meta says it’s “small enough to run on a Mac or PC with a single consumer GPU.” That puts it within reach of a well-specced Mac mini, which is the machine I price out below. The release also shows where Meta draws its line: as of October 22, 2026, its most capable models stay behind an API or a subscription, while open weights arrive one tier down.

Key takeaways

No user cap and no EU carve-out. Glimmer’s license file is the plain Apache 2.0 text, with none of the Llama 4 restrictions. (details)
A separate usage policy still exists. Whether it binds you on top of Apache 2.0 is an open question. (details)
Open weights that fall short of “Open Source AI.” By OSI’s definition, the missing training code and data details keep it out. (details)
A 24 GB Mac mini is the tight minimum and 32 GB is comfortable. That’s my estimate from Meta’s 4-bit builds, which run 17 to 23 GB in memory. (details)
Muse Spark Open is the bigger sibling with strings attached. It has a custom license, a gated download, and needs a Mac Studio cluster to run. (details)
Scores are mixed. Glimmer leads on some agent benchmarks and posts an 82% hallucination rate on Artificial Analysis’s test. (details)

What Muse Glimmer is, and how it differs from Muse Spark and the Meta Muse app

Three Meta names show up in this post, so here’s how they fit together. Meta Muse is the consumer agent app I covered in my October 3 explainer. Muse Spark is the large model that powers it, available through Meta’s API. Muse Glimmer is the small model you download, which Meta’s model card says was distilled from Muse Spark (distillation means training a small model to imitate a bigger one’s answers).

The card lists 29.6 billion parameters, which Meta rounds to 30B, including a vision encoder. It takes text and images in and produces text out, with a 131K-token context window, and Meta built it for tool use, multi-step tasks, and reading screenshots and documents. Meta’s developer docs say there’s “no API key and no gated access,” and the Hugging Face repo metadata confirms it. A technical report followed on October 17.

Glimmer sits at the bottom of a lineup that gets more closed as it gets bigger. Muse Spark 1.3 (September 2) and Muse Wonder (October 12) are available only through Meta’s API and its Muse Code tool, and Wonder Full only with a Muse Code subscription. Meta calls Wonder “our most capable models to date.”

Meta’s model docs list just two open-weight lines, Glimmer and Muse Spark Open, and those are the only two in Meta’s Hugging Face organization as of October 22. Asked about open weights for Wonder, AI at Meta replied that those models “are too big for many people to use, in addition to potential safety concerns.”

Apache 2.0 vs the Llama 4 Community License

Apache 2.0 grants a “perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable” right to copy, modify, and redistribute the work, plus a patent license. Glimmer’s LICENSE file is that standard text, unmodified, with no mention of user counts or Europe.

The Llama 4 Community License reads differently. Any company with more than 700 million monthly active users has to ask Meta for a license, “which Meta may grant to you in its sole discretion.” Its acceptable use policy withholds the multimodal models from individuals and companies based in the European Union (end users of products built with them are exempt), and that policy is written into the license itself.

	Muse Glimmer	Llama 4
License	Apache 2.0	Llama 4 Community License
User cap	None	Over 700M monthly users need Meta’s permission
EU limit	None	Multimodal rights not granted to EU-based people or companies
Separate use policy	Yes, wording “You should not use”	Yes, written into the license, “You agree you will not use”

Google’s Gemma 4, Alibaba’s Qwen3.6-27B, and OpenAI’s gpt-oss all ship under Apache 2.0 already, so Meta is catching up to the field rather than leading it. Artificial Analysis gives Glimmer a 44 on its Openness Index, above Llama 4 Maverick, crediting the license and Meta’s methodology disclosure.

Does Meta’s usage policy still apply to Muse Glimmer?

I’m not a lawyer, so treat this as a reader’s note and not advice. The model card links a Usage Policy that “applies to your access or use of Muse Glimmer.” It lists prohibited uses (military and nuclear work, critical infrastructure, malware, disinformation, and “representing that the use of Muse Glimmer or outputs are human generated”) and says the model “is not intended for individuals under the age of 18.”

The Apache license file never mentions that policy, and the policy says “You should not use” where Llama 4’s said “You agree you will not use.” Whether a standalone policy adds enforceable conditions on top of Apache 2.0 is a question I can’t answer. Stefano Maffulli, writing on his personal blog, landed in the same place: “What that policy is or does, I’m not sure. Lawyers?” OpenAI’s gpt-oss pairs Apache 2.0 with a one-line policy of its own, so the pattern isn’t unique to Meta.

Muse Glimmer vs Muse Spark Open

On October 15, Meta released Muse Spark Open, which its developer blog calls “the open weights version of Muse Spark” and which Mark Zuckerberg had promised on Glimmer’s launch day. The model card lists 816 billion total parameters with 42 billion active, a mixture-of-experts design (only part of the model works on each word), a 1-million-token context, and “additional safety post-training.” Meta doesn’t say whether it otherwise matches Spark 1.2.

The strings are in the license and the download. Spark Open uses a custom Muse Model License, and the full text makes the usage policy binding (“hereby incorporated into this license”), requires a “Built with Muse Spark Open” notice on derivatives, and requires model-as-a-service sellers with more than $20 million in revenue to sign a separate agreement with Meta. Meta calls these “limited commercial terms, which allow many developers to use the model for free.”

Neither Spark Open nor Glimmer carries Llama 4’s user cap or EU carve-out. The official download needs manual approval, although community conversions are ungated.

	Muse Glimmer	Muse Spark Open
License	Apache 2.0	Custom Muse Model License, usage policy binding
Download	Ungated	Manual approval (name, birth date, country, job title)
Size	29.6B dense; 16.8 GB at 4-bit	816B total, 42B active; smallest build about 176 GB
Hardware	24 to 32 GB Mac mini	Two M5 Ultra Mac Studios (Meta’s reference setup)

The smallest Spark Open community build I found is about 176 GB, and the largest Mac mini Apple sells has 64 GB of memory, so no Mac mini can run it. Meta’s reference local setup is two M5 Ultra Mac Studios, and Ollama offers Spark Open only as a cloud model.

Is Muse Glimmer open source by OSI’s definition?

The Open Source Initiative’s Open Source AI Definition requires the freedom to use, study, modify, and share a system “for any purpose.” It also requires three things: enough data information that “a skilled person can build a substantially equivalent system,” the complete source code used to train and run the system, and the parameters (the weights). It adds that “Open Source weights” must include the data information and code used to produce them.

By my reading, Glimmer meets one of the three. The weights are out under an OSI-approved license. I found no training or data-processing code (Meta points to PyTorch’s TorchTitan for fine-tuning, which is a different thing).

The data description stays high-level: “publicly available data, data provided by third parties and information from Meta’s products and services.” A third-party-hosted copy of Meta’s EU training-data summary adds that the text ran past 10 trillion tokens and included public Facebook and Instagram posts, with no dataset listing. So Glimmer is open weights under an open-source license, and it falls short of Open Source AI as OSI defines it.

OSI hasn’t said anything about Glimmer by name as of October 22, though its September 16 post, Open weights are good, open source is better, makes the general case: “Open-weights allow you to exercise some freedoms, but not all.”

How good is Muse Glimmer?

Meta’s own numbers, run at high reasoning effort against Gemma 4 31B and Qwen3.6-27B, favor Glimmer on agent tasks. It leads on MCP Atlas (75.5 vs 54.2 and 62.5) and DeepSearch QA (74.6 vs 61.7 and 71.1). Qwen wins on SWE-Bench Verified, TerminalBench, and OSWorld-Verified, where it posts 75.6 to Glimmer’s 65.9.

Meta’s methodology note says competitor numbers are “the most favorable result between self-reported scores or our internal reproductions.” That’s a reasonable way to pick the rival scores, though Meta still chose the tests and ran the comparison.

Artificial Analysis, which had pre-release access, scored Glimmer 35 on its Intelligence Index at launch, behind Qwen3.6 27B at 38 and ahead of Gemma 4 31B at 30 and Llama 4 Maverick at 14. Its hallucination test is the sore spot: an 82% rate against Qwen’s 49%. The current model page shows a 17 on index version 4.3.2, which uses a different scale, so the drop isn’t a regression.

How to run Muse Glimmer on a Mac mini

Apple sells two Mac mini chips. The M6 model starts at $899 with 16, 24, or 32 GB of memory, and the M5 Pro model starts at $1,699 with 24, 48, or 64 GB. Memory is the number that matters, because a model has to fit in it to run (I covered why in my post on Apple Intelligence’s on-device limits).

Meta’s official 4-bit builds are 16.8 GB and 19.7 GB, with totals of roughly 17 to 23 GB once you add the vision file and the optional speed-up drafter. Meta’s target is 24 GB, while LM Studio says “you need at least 26 GB of RAM.” One tester on a 24 GB M4 Pro laptop found the 17 GB build worked for text and ran out of memory with vision at a longer context, calling 24 GB “a minimum, not a comfortable all-features envelope.”

Here’s how I’d buy for running an agentic model like Glimmer. Amazon links use my affiliate tag where a clean listing exists; higher-memory configs often ship configure-to-order from Apple only.

Apple Mac mini (M6, 16GB) (affiliate link). Skip this for Glimmer. Even the small community builds are a squeeze, and an agentic loop with tools or images won’t fit comfortably.
Apple Mac mini (M6, 24GB) (affiliate link). It runs text-only chats with short context, with little room to spare. It’s fine for a quick try, but expect to turn vision and long chats off.
Apple Mac mini (M6, 32GB). This one should be comfortable at 4-bit, images included, and it’s the config I’d pick for daily agent use. As of October 23, 2026 I couldn’t find a clean Amazon listing for this config, so buy it configured on Apple.
Apple Mac mini (M5 Pro, 24GB) (affiliate link). It has the same memory squeeze as the 24GB M6, but the Pro’s higher memory bandwidth should feel snappier on the same short text chats.
Apple Mac mini (M5 Pro, 48GB or 64GB). These have room for long context, the optional speed-up drafter, or 8-bit builds. Full precision still won’t fit even at 64GB. Configure these on Apple’s Mac mini store; Amazon listings for 48GB and 64GB weren’t verified for this draft.

Mac mini (M6 from $899, M5 Pro from $1,699)	Memory	Bandwidth	Fit (my estimate)
M6	16 GB	153 GB/s	Not practical; only small community builds fit
M6	24 GB	170 GB/s	Text only, short conversations, tight
M6	32 GB	170 GB/s	Comfortable at 4-bit, including images
M5 Pro	24 GB	307 GB/s	Same tightness as the M6, faster
M5 Pro	48 GB	307 GB/s	Long context, drafter, or 8-bit
M5 Pro	64 GB	307 GB/s	8-bit comfortable; full-precision still won’t fit

The fit column is my estimate from file sizes and the tester reports above, since nobody has published Mac mini runs for each configuration. The easiest path is Ollama: install it and type ollama run muse-glimmer, which pulls an 18 GB build (a 30b-mlx tag uses Apple’s MLX engine and handles images). LM Studio works too if you’d rather click than type, and llama.cpp needs build b10353 or later.

One Hacker News user reported running it “on my old MacMini 32G using Ollama” with “good results,” adding “with the caveat that everything runs slowly” and “remember to increase the context size!” A measured run on an M4 Pro with 48 GB hit 14.7 tokens per second at 4-bit. I haven’t seen measurements on the M6 or M5 Pro.

How people are reacting to Muse Glimmer

The launch thread on Hacker News drew 1,209 points and 638 comments, and the mood was positive with reservations. People welcomed the license and the speed. A Hugging Face commenter wrote, “This model is so damn fast,” while another in the same thread said it “reasons SO heavily about policy adherence that it is incoherent.”

One commenter argued that “open weight” models “are still just binary blobs that are completely inscrutable.” Another pointed out that it “still needs 32-64GB memory to run it locally.” The most-replied comment held that “this is open weights because Meta couldn’t monetize it in any other way.”

In October, two months after launch, one user called it “by far the most usable model on my 32GB M1 Max,” while another kept Qwen because Glimmer “will vacillate wildly if pushed.” Each of those is one person’s run on one kind of task, so I’d read them as anecdotes until someone publishes a broader test.

What to watch next with Muse Glimmer

Meta promised that its DFlash drafter, a speed-up add-on that guesses several words at a time, would make Glimmer 1.5x to 3.1x faster with “identical output quality.” Independent testers haven’t matched that. Lights Out Labs found that outputs differ with and without the drafter, the 24 GB tester measured a run that was 33.5% slower with it, and one small X test on twelve accounting questions reported accuracy dropping to 75%. Whether Meta closes that gap in a later build is the first thing I’d watch.

The second is whether Apache 2.0 sticks. Spark became Spark Open in two months, under a custom license with a gate, and Wonder is API and subscription only today for reasons Meta describes as size and safety. Meta’s next open release will show whether the Glimmer license was a change in policy or a one-time choice for the one model it could afford to give away.

If you’ve run Glimmer on a Mac mini, I’d like to hear which configuration and what you handed it first.

Frequently asked questions

Is Muse Glimmer open source? Only the license is open source, by my reading. Meta released the weights under Apache 2.0, an OSI-approved open-source license. The model as a whole doesn’t meet OSI’s Open Source AI Definition because Meta hasn’t released the training code or a detailed description of the training data. “Open weights” is the accurate label.

Can I use Muse Glimmer for commercial projects? Apache 2.0 allows commercial use, modification, and redistribution without asking Meta or paying anything, and the license file has no user cap. Meta also publishes a separate Usage Policy with prohibited uses and an 18-plus note. Whether that policy is binding on top of Apache 2.0 is unclear, so check with a lawyer for anything serious.

How is Muse Glimmer’s license different from Llama’s? The Llama 4 Community License requires companies with more than 700 million monthly users to get Meta’s permission, and its use policy withholds multimodal rights from people and companies based in the EU. Glimmer’s Apache 2.0 license has neither restriction and adds a patent grant.

How much RAM do I need to run Muse Glimmer, and can a Mac mini do it? Meta’s 4-bit builds take 17 to 23 GB of memory depending on whether you add images and the speed-up drafter, and LM Studio says you need at least 26 GB. A 24 GB Mac mini works for text with short conversations, and 32 GB is comfortable, in my estimate. Apple’s M6 Mac mini starts at $899 and goes up to 32 GB, and the M5 Pro model starts at $1,699 with up to 64 GB.

Is Muse Glimmer the same thing as Meta Muse? No, they’re two different products. Meta Muse is the consumer agent app that runs on Meta’s servers, powered by the large Muse Spark model. Muse Glimmer is a 30-billion-parameter model distilled from Muse Spark that you download and run yourself. You can’t swap Glimmer into the Muse app.

What’s the difference between Muse Glimmer and Muse Spark Open? Muse Spark Open is Meta’s 816-billion-parameter open-weights model, released October 15, 2026 under a custom Muse Model License with a binding usage policy and a gated download. Glimmer is 30 billion parameters under plain Apache 2.0 with no gate. Spark Open’s smallest build is about 176 GB, so it needs Mac Studio-class hardware, while Glimmer fits a 24 to 32 GB Mac mini.

Does Muse Glimmer work offline? Yes, it runs fully offline once downloaded. Meta’s announcement describes it running “with or without an internet connection,” and the developer docs say there’s “no API key and no gated access.” Once you’ve downloaded the weights through Ollama or LM Studio, nothing leaves your machine.