Muse Glimmer: What's New, How to Run It, and Alternatives (2026)
Muse Glimmer is a 30-billion-parameter open-weight model from Meta Superintelligence Labs, released on August 10, 2026 under the Apache 2.0 license. Meta's own headline for it is direct: "An Open Agentic Model That Runs on Your Device."
That last phrase is the whole point. This is not a chat model that happens to call tools. It is a model sized so an agent loop can run on hardware you already own.
This guide covers what the release actually contains and how to run it. It also covers what the official announcement leaves out, and where the model sits against the other open-weight options.
What is Muse Glimmer?
Muse Glimmer is a single model with a stated focus on agentic work. Meta lists the capabilities as end-to-end task completion, reliable tool use, multi-step reasoning and failure recovery.
That fourth item deserves attention. Failure recovery is what separates a demo from a task that finishes. An agent that cannot notice its own bad step will loop or stop.
The model takes text and images. Meta describes the image path as a dedicated perception encoder rather than a bolted-on adapter.
It also lists scaffold compatibility, controllable effort and support for more than 100 languages. Scaffold compatibility matters because you will likely run it inside someone else's agent harness.
Controllable effort is the one that shows up in daily use. It lets you spend more compute on a hard step and less on a trivial one. That is how a long task stays affordable on hardware you paid for once.
What's new in the August 2026 release
Three things are genuinely new here, and they are easy to conflate.
The size and the license together. A 30B model under Apache 2.0 is permissive enough for commercial use without a bespoke agreement. Meta calls it a "permissive Apache 2.0 license."
The device framing. Meta's own wording is that the model is "small enough to run on a Mac or PC with a single consumer GPU." The announcement names the machines it was validated on.
The quantization claim. Meta states that quantization "introduces minimal to no degradation on agentic tasks." That is a stronger claim than the usual quality note, because agentic tasks are where quantization damage normally shows up first.
Weights are published on Hugging Face. Meta's line is that the model "is available now, and you can download the weights on Hugging Face."
What Meta published about benchmarks, and what it left out
Here is the part most coverage skipped. The announcement does not print benchmark numbers.
It says the model was compared with Gemma4-31B and Qwen3.6-27B across agentic, coding, multimodal, safety and reasoning benchmarks. For the figures themselves, it points to a separate report.
Now look at the comparison set. Qwen3.6-27B is not the current Qwen model in that size class. Qwen3.8-27B shipped on August 14, four days after this announcement.
So the peer table was one generation behind within a week of publication. That is nobody's fault. It is simply what the release cadence does to any published comparison right now.
The practical lesson is to treat a vendor peer table as a snapshot with a date, not as a ranking. If you care about the answer, run your own file through both.
There is a second gap worth naming. The announcement gives no context window figure.
For agent work on documents and spreadsheets, that number decides what fits in one pass. Its absence is the single most relevant unknown for anyone planning to feed the model real files.
How to run Muse Glimmer
Meta lists the runtimes directly, which makes this unusually easy to plan.
| Route | What it is | Best when |
|---|---|---|
| Hugging Face | Official weights download | You want the unmodified checkpoint |
| llama.cpp | Cross-platform local runtime | You need broad hardware support |
| MLX | Apple-silicon runtime | You are on a Mac |
| ExecuTorch | On-device deployment path | You are shipping to devices |
| vLLM / SGLang | Serving stacks | You are hosting it for a team |
| Ollama / LM Studio | Packaged local apps | You want the shortest setup |
| Together AI / Fireworks AI / OpenRouter | Hosted API partners | You do not want to run it at all |
The last row is worth calling out. An open-weight model does not oblige you to host it. Several partners serve it, so you can test the model before you buy hardware for it.
What it takes to run it well
Meta names the machines it validated: MacBook M4-Max, M5-Max and the RTX-5090. Those are the reference points, not the minimum spec.
It also publishes speculative decoding gains, which is the most concrete performance detail in the release.
| Hardware | Decode speed increase with speculative decoding |
|---|---|
| RTX 5090 | 3.1x |
| M5 Max | 1.8x |
| M4 Max | 1.5x |
Read those as a hardware ladder. The same technique pays back roughly twice as much on the desktop GPU as on the older laptop.
One caveat belongs with that table. Speculative decoding needs a second, smaller model running alongside the main one, and that costs memory. Treat the figures as an upper bound rather than a free upgrade.
Be honest with yourself about the cost side too. "Free weights" and "free to run" are different statements. The machines above are not cheap, and the electricity is yours.
Turning a local model's output into something you can send
Running the model is one job. Producing the thing your colleague asked for is another.
A local agent can read your export and reason about it. It still leaves you assembling the chart, the table and the wording into a document someone will actually open.
That gap is where a hosted agent earns its place. Powerdrill Bloom takes the file and returns the deliverable. You hand over slides, a spreadsheet or a written summary without wiring a pipeline first. Its Pro plan is $13.27 per month billed annually, which is one comparison point against buying a GPU.
The trade is real in both directions. Local means your data never leaves the machine, and you own the whole stack. Hosted means no setup and no hardware, and you accept that the files go to a service.
There is a third shape in this market worth knowing about. Our note on GenOffice covers an open, self-hosted office suite rather than a bare model.
Alternatives worth comparing
Qwen3.8-27B, released August 14, 2026, is the most direct comparison. Its model card lists 262,144 tokens of native context, text, image and video input, and an Apache 2.0 license. It is also the model that replaced the one in Meta's peer table.
Gemma4-31B is the other model Meta names, so it is already positioned as a peer in the same size class.
Hosted API models are the honest alternative for most teams. If you never wanted to manage a runtime, a metered API removes the hardware question entirely. Several of the partners Meta lists serve this model that way, so the open licence does not force you into hosting.
Agent platforms are a different layer, not a competing model. If your goal is a finished report, the model is one component. Our explainers on what a general-purpose data agent is and on MCP cover how those pieces connect.
Conclusion
Muse Glimmer is a 30B Apache 2.0 model built for agent loops on your own machine. Meta validated it on high-end laptops and a consumer GPU. The device framing is the news here, not the leaderboard position.
Two caveats should shape your expectations. The announcement publishes no context window, and its peer comparison was outdated within four days.
As of August 2026 those are the facts on the official page. Your real goal may be a chart, a deck or a written report from a file you already have. Test that whole path before you commit to any model. Our roundup of AI data agents for business teams and the auto insights page are reasonable places to start.
Frequently asked questions
Is Muse Glimmer free?
The weights are published under the Apache 2.0 license, so downloading and using them carries no license fee. Running the model still costs you hardware and power, or a hosted provider's fee.
What hardware do I need?
Meta describes it as small enough for a Mac or PC with a single consumer GPU. The machines named in the announcement are the MacBook M4-Max, the M5-Max and the RTX-5090.
How big is the context window?
Meta's announcement does not state one. If your work depends on feeding long documents in a single pass, treat that as an open question until an official figure appears.
How does it compare with Qwen3.8?
Meta's published comparison uses Qwen3.6-27B, not Qwen3.8-27B, which was released four days later. Any current comparison has to be run yourself or taken from a dated third-party evaluation.
Can it produce slides or a report on its own?
The model handles reasoning and tool use. Turning that into a formatted document depends on the agent scaffold you run it inside, not on the model alone.