Muse Wonder vs Muse Glimmer: One Is a Rate Card, the Other Is a File

Meta Superintelligence Labs now sells two models that share a surname and almost nothing else. Muse Wonder, released October 12, 2026, is a proprietary frontier model at $4.00 per million input tokens and $20.00 per million output with a 1,048,576-token window and an Artificial Analysis Intelligence Index of 52. Muse Glimmer, released August 10, 2026, is a 30-billion-parameter Apache 2.0 model that Meta ships already quantized to 4-bit so it fits under 20 GB of VRAM, runs with the network unplugged, and scores 17 on the same index. That is a 35-point gap, and it is not the interesting number. The interesting number is 20 GB: one of these two you can hold in your hand, and the other one you can only rent.

Two products, not two tiers

It is tempting to read Muse Glimmer as a cut-down Muse Wonder. The provenance runs the other way. Glimmer was trained by logit distillation from Meta's larger Muse Spark teacher, then fine-tuned on long-context agentic data and reinforced for tool use — it is a small model built to imitate a big one, not a big one with features removed. Muse Wonder came out of the same lab eight weeks later on a different track entirely: 1M context, speech and video as first-class inputs, mandatory reasoning on an effort ladder running from minimal to max, and a subscription-only twin checkpoint behind it.

• Weights — Muse Glimmer Apache 2.0, downloadable, shipped at 4-bit; Muse Wonder proprietary, no checkpoint published

• Parameters — Muse Glimmer 30 billion, under 20 GB quantized; Muse Wonder undisclosed

• Where it runs — Muse Glimmer on your own GPU, offline; Muse Wonder on Meta's infrastructure over a network

• Context window — Muse Glimmer 131,072 tokens, extendable to 262,144; Muse Wonder 1,048,576 tokens

• Modal input — Muse Glimmer text and image; Muse Wonder text, image, speech and video

• Price — Muse Glimmer no per-token charge if you self-host, $0.325 in and $1.35 out per million if you rent it; Muse Wonder $4.00 in and $20.00 out

• Output ceiling — Muse Wonder 943,718 tokens in a single response; Muse Glimmer not documented at that scale

• Knowledge cutoff — Muse Glimmer January 2026; Muse Wonder not published

• Absorbable change one of them cannot absorb — Muse Glimmer can be pinned, forked and fine-tuned; Muse Wonder can be deprecated

Meta tested Glimmer on an RTX 5090 and on Apple M4 Max and M5 Max silicon, which is the practical definition of what "runs locally" means here — the full-precision footprint is over 55 GB, so the 4-bit ship is the whole reason the model is deployable at all. Nothing about Muse Wonder works this way. You reach it over an API, at a rate, for as long as Meta chooses to serve it.

What the 35-point gap is made of

A gap that size usually means one model is simply worse. Here it mostly means the two models were built for different jobs, and the sub-scores show it rather than hiding it. All figures below are Artificial Analysis measurements on revision v4.3.2 unless marked otherwise.

• Intelligence Index — Muse Wonder (max) 52, ninth of 219 models; Muse Glimmer (high) 17, fifteenth of 142

• Long-context recall — Muse Glimmer 83.3%, ahead of Claude Sonnet 5.5's 82.7% in the same run

• Terminal-Bench 4.0 — Muse Glimmer 0.5%

• Automation benchmark — Muse Glimmer 0.068

• Humanity's Last Exam — Muse Glimmer 22.0%

• Output speed — Muse Glimmer 143.8 tokens per second; Muse Wonder not published on the same harness

• Cost per index task — Muse Glimmer about $0.06 hosted; Muse Wonder about $4.52

The long-context recall line is the one worth pausing on. A 30-billion-parameter model quantized into 20 GB matching a brand-new frontier model at finding a fact in a very long input is not a rounding artefact — it is what the training recipe was aimed at, and it is the strongest claim anyone can make for Glimmer. The Terminal-Bench line is the other end: half a percent means the model is not completing agentic software tasks, and no saving on hardware makes a task that never finishes cheaper. A peer-reviewed multi-agent orchestration study that included Muse-Glimmer-30B found it recovered none of its candidates, which is the same finding from a different direction.

So the shape is legible. Glimmer reads well, answers fast, runs anywhere, and cannot execute. Wonder reads a larger input, executes, and costs roughly seventy-five times as much per index task when you rent the small one instead of owning it.

The $0.06 number is a trap

That seventy-five-fold figure deserves immediate disqualification. The $0.06 per index task is what Artificial Analysis spent calling Muse Glimmer through a hosted endpoint at $0.325 and $1.35 per million — it prices the economics of renting the model, not of running it. Self-hosted, the per-token bill disappears and is replaced by a GPU you already own or have to buy, plus the engineering time to serve it. Quoting $0.06 against $4.52 as though it were a like-for-like comparison is the single easiest mistake to make on this page.

The honest version is that the two models are not on the same axis at all. Muse Wonder is priced by the token and its cost scales with volume. Muse Glimmer is priced by the accelerator and its cost does not scale with volume — which makes it expensive at ten requests a day and remarkably cheap at ten million.

Offline is a requirement, not a discount

For one class of deployment this comparison is not a comparison. A request that cannot leave the device, a factory floor or field site with no connectivity, an air-gapped environment where an API call is itself a compliance violation, a latency budget where a round trip is already too much — none of those are solved by a better model behind a network. Apache 2.0 weights that fit under 20 GB and run unplugged are the entire answer, and Muse Wonder does not participate in the question at any price.

That is also the one place where the licence stops being a footnote. Meta's own launch material for Muse Wonder says nothing about weights, and none are coming; the Full checkpoint behind the Muse Code subscription is likewise closed. Glimmer is a file. You can fork it, fine-tune it, pin the revision, and keep serving it after the vendor loses interest — which for a model released two months ago by a lab that has already shipped two further generations is not a hypothetical concern.

Holding all three, which is the actual answer

Most teams that read this far are not choosing one. They want a small model on their own hardware for the cheap, private, high-volume half of the workload and a frontier model over an API for the hard half, and the awkward part is that those two halves normally mean two entirely separate stacks.

One of the three Meta models in this family is a route we serve. meta/muse-spark-1.2 — the larger teacher Glimmer was distilled from — sits in the OrcaRouter catalogue at $1.25 input and $4.25 output per million tokens across a 1,048,576-token window with $0.15 cache reads, which makes it the middle rung: considerably more capable than the 30B student, a fraction of Muse Wonder's rate, and reachable on the same key as everything else. It is also carrying 42.8 million tokens of traffic a week through this catalogue, which is the useful signal for anyone wondering whether the Spark line does real work.

Two honesty notes. Muse Glimmer itself is not one of our routes — there is no muse-glimmer entry in the catalogue, and the capture below is Muse Spark 1.2's page rather than Glimmer's, because Glimmer's whole point is that you run it yourself and never route it anywhere. Muse Wonder is not ours either; Meta's Model API and Muse Code are where it runs.

What the single-endpoint arrangement buys is the hosted half of that split. Provider list prices pass through at 0% markup, so a vendor repricing shows up the same day; automatic failover covers an upstream provider degrading mid-run; and the routing DSL lets you compose a cheap model and an expensive one into a single call so that the frontier model is invoked only where the cheap one's confidence drops. For a family with a 30B model at one end and a $20-per-million-output model at the other, that composition is the whole point — most traffic never needs the expensive end.

Choosing

• Muse Glimmer when the tokens cannot leave the machine, when there is no network, when you need something you can pin and keep, or when volume is high enough that a per-token bill becomes the dominant cost. Within that envelope it is fast, concise and strong at long-context reading, and the composite score understates it for exactly those jobs.

• Muse Glimmer for high-volume extraction and retrieval over long documents, where the 83.3% long-context recall is the number that matters and the 0.5% on Terminal-Bench never comes up.

• Muse Wonder when the task has to be executed — agents, computer use, terminal work — or when the input is speech or video, or when a single response needs to run past 128,000 tokens. A 0.5% execution score is not a budget option; it is a non-participant.

• Neither, if you are shopping in the middle. Muse Spark 1.2 is the rung between them, it is the model Glimmer was distilled from, and it costs a fifth of what Muse Wonder charges for capability somewhere in the gap.

The 35-point index gap is real and it is also the least useful thing on this page. These two models are not adjacent products with a price step between them; they are two answers to questions that only overlap on the days you happen to need both. The decision is where the tokens run, and the second decision — which is the one people forget — is that a lab shipping this far apart at both ends of the range will keep moving, so the model you cannot pin is the model you should be slowest to build on.