Meta Is Back in Open Weights: What Muse Glimmer 30B Actually Runs On

On 10 August 2026 Meta released Muse Glimmer, a 30 billion parameter model, under Apache 2.0. Weights are on Hugging Face, GGUF conversions appeared within hours, and vLLM announced day-zero support with the phrase “Meta is back in open source”.

Four days earlier we published a post arguing that Meta had gone closed. Well.

That post, what Muse Spark meant for open weights, described Muse Spark 1.2 shipping behind an API with no open Llama flagship in over a year, and concluded that the open weights torch had passed to other labs. The first half was accurate. The conclusion did not survive contact with the following week.

What Shipped

Muse Glimmer is a 30 billion parameter dense model, not a mixture of experts, so every parameter activates on every token rather than routing through a subset. Meta describes it as built for agentic work: multi-step tasks, tool use, long trajectories. A perception encoder lets it read images, screenshots, and documents alongside text.

The licensing is the headline. Apache 2.0 with no usage restrictions, not a Meta-branded community license with a monthly-active-user threshold buried in clause seven. Meta released BF16 weights, 4-bit quantized variants, and the speculative decoding drafter under the same terms. The Muse Glimmer model card on Hugging Face is the thing to read rather than coverage of it, since license claims in secondary reporting are wrong often enough to matter. Open weights model licenses explained covers why Apache 2.0 removes a category of legal review that community licenses create.

“Runs on One Consumer GPU” Deserves an Asterisk

Every launch post said it runs on a single consumer GPU. This is true and it is doing a lot of work.

The community 4-bit GGUF quantizations of Muse Glimmer land in the region of 17 to 20 GB depending on the variant, before you add context. That is a card with 24 GB of VRAM, or a Mac with a large unified memory pool, or a patient CPU-only setup with enough system RAM and no expectations about speed.

Now think about the machine your employer issued you. Sixteen gigabytes of system RAM, integrated graphics, a docking station, and a mandatory endpoint agent taking a slice of all of it. “One consumer GPU” describes a gaming desktop or an enthusiast’s workstation. It does not describe the fleet laptop most people are reading this on.

This is not Meta overselling anything. Getting a 30B multimodal agentic model onto a single card is a real achievement. The phrase has simply drifted, and the gap between “runs locally” and “runs on your work laptop” is where most disappointment with local AI comes from. The sizing procedure in how to choose an open weights model for your laptop starts from the memory you actually have free, which is the only order of operations that works.

Agentic Capability Is Not Rewriting Capability

Set the hardware aside and ask a different question: if you could run it, would it write a better email?

Almost certainly not in any way you would notice. A rewrite has no hidden depth. You give the model a paragraph and a constraint, it returns a paragraph. No plan to form, no tool to call, no state to carry across twelve steps. Every capability Muse Glimmer was distilled to be good at is capability a tone change does not exercise.

The failure modes that ruin rewrites are also unchanged by scale. Verbosity, invented specifics, softened commitments, a cheerful closing sentence you did not ask for. A 30B agentic model does all of those with more fluency than a 4B instruction model, which is the wrong direction. The fuller argument is in small language models beat big ones for rewriting.

For reference, Wrivio’s Local engine ships Apache 2.0 Qwen3 models at 1.7B (about 1.1 GB on disk) and 4B (about 2.4 GB), sized so they run while you have Outlook, Teams, and thirty browser tabs open. Different goal, different model.

Where Muse Glimmer Genuinely Matters

It matters upstream of you.

A permissively licensed 30B multimodal model that a small company can host on its own hardware changes what private deployments cost. It gives distillation and fine-tuning work a legally clean base. Over the next year some of that flows down into smaller models you can actually run, which is how every previous open release has worked.

It also matters as a signal about Meta, and signals from Meta on this subject have been unreliable in both directions. One Apache 2.0 release is not a strategy, and Muse Spark remains API-only. Be pleased, and do not build a plan around Meta’s next decision.

Judging a Release Without Downloading It

The instinct on release day is to fetch the weights and form an opinion afterwards. Reverse it. Read the model card, find the memory requirement, compare to your free RAM, and stop there if it does not fit.

Before:

Just downloaded the new Meta model, going to see how it does. Might be a big upgrade for us.

After:

Muse Glimmer 30B is Apache 2.0, released 10 August 2026. The 4-bit builds are roughly 17 to 20 GB, so it needs a 24 GB GPU or a large unified-memory Mac. Our laptops have 16 GB total, so it is not a candidate for local use here. Worth watching if smaller distilled versions appear.

The second took no longer to write and answers the question the reader actually had.

A Wrivio Context for summarising a model release for colleagues could say:

Rewrite this as a short factual note for a technical team. Lead with the licence, the parameter count, and the memory requirement. Keep every name, date, figure, and licence term exactly as written. Do not add benchmark claims, opinions about quality, or recommendations that are not in the original.

Press Ctrl+Shift+Space, paste your notes, and check the diff. Model posts are exactly where an AI rewrite likes to insert a plausible benchmark number that nobody stated, which is why why a good AI rewriter never invents facts is the habit to build here.

Common Questions

What licence is Muse Glimmer released under?

Apache 2.0, covering the BF16 weights, the 4-bit quantized variants, the speculative decoding drafter, and the perception encoder. Confirm the terms on the Hugging Face model card before relying on this for a commercial deployment.

How much VRAM does Muse Glimmer 30B need?

Community 4-bit GGUF builds are roughly 17 to 20 GB before context, which in practice means a 24 GB GPU or a Mac with a large unified memory pool. Full precision is far larger.

Can I run it on a normal work laptop?

Not comfortably. A typical 16 GB corporate laptop with integrated graphics does not have the headroom, and running it on CPU alone would be too slow for interactive use.

Is a 30B model better than a 4B model for rewriting emails?

Not in a way most people would notice. Rewriting rewards precise instruction-following rather than reasoning depth, and that ability appears in much smaller models.

Does this mean Meta is committed to open weights again?

One Apache 2.0 release is evidence, not a commitment. Meta’s frontier Muse Spark line remains API-only as of August 2026, so treat the direction as unsettled.

Download Wrivio for Windows to run a rewrite locally on a model sized for the laptop you actually have, with no network call during the rewrite.