Muse Wonder Full vs Muse Glimmer: the Best Score in the Table Is Also the Worst Number in It

Muse Wonder Full and Muse Glimmer are the two ends of Meta Superintelligence Labs' published range. Muse Wonder Full is a proprietary frontier checkpoint announced October 12, 2026, sold as part of the Muse Code subscription and scored at 59.1 on Meta's own index. Muse Glimmer is a 30-billion-parameter Apache 2.0 model released August 10, 2026, shipped already quantized to 4-bit so it fits under 20 GB of VRAM, scoring 17 on the independent Artificial Analysis Intelligence Index. Between them sit 42 points of composite score, a licence, a download, and a $50-a-month door — and one row in Meta's own table where the expensive one scores 13.1.

That row is Harvey LAB-AA, Meta's legal agent benchmark. Muse Wonder Full reports 13.1 on it. Muse Wonder, the callable checkpoint from the same release, reports 54.3, GPT-6 Astra 54.5 and Claude Opus 5.5 55.2. A model that leads the composite index by four points finishes last on a job-shaped suite by a margin of 41. That is the first thing to understand about this pairing: the composite score does not describe the work.

Two products, one surname, and a third that gets confused with both

• Weights — Muse Glimmer Apache 2.0, downloadable, shipped at 4-bit; Muse Wonder Full proprietary, no checkpoint published, and none coming.

• Where it runs — Muse Glimmer on your own accelerator, offline; Muse Wonder Full on Meta's infrastructure inside Muse Code.

• How you buy it — Muse Glimmer no per-token charge if you self-host, or $0.325 in and $1.35 out per million through a hosted endpoint; Muse Wonder Full no per-token price at all, on published plans of $50 a month for Power Usage and $200 a month for Maximum Usage.

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

• Context window — Muse Glimmer 131,072 tokens, extendable to 262,144; Muse Wonder Full described alongside the series at 1,048,576 tokens.

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

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

• What can be done to it — Muse Glimmer can be pinned, forked and fine-tuned; Muse Wonder Full 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 "runs locally" here — full precision is over 55 GB, so the 4-bit ship is the entire reason the model is deployable. Nothing about Muse Wonder Full works this way. It is reached over a subscription, at a plan price, for as long as Meta serves it.

One more distinction worth stating plainly, because the naming quietly collapses it: Muse Wonder Full is not the top of the Muse Wonder series in any general sense, and it is not a bigger Glimmer. It is a separate training target announced in the same breath as Muse Wonder, and the only one of the three you cannot buy per token, self-host, or evaluate without a purchase order.

The licence is the axis nobody prices

This comparison is usually framed as capability against cost. The framing that actually predicts what you can do is capability against control, and the two models sit at opposite poles of it.

Everything about Muse Glimmer's design assumes an owner rather than a customer. The weights are Apache 2.0, the quantization is done for you, the footprint is under 20 GB because that is the number that makes a single consumer card sufficient, and the knowledge cutoff is published so you know what it does not know. You can fork it, fine-tune it, pin a revision, and keep serving it after the vendor loses interest. For a model released by a lab that has already shipped two further generations since, that last property is not hypothetical.

Muse Wonder Full has the inverse profile. There is no artifact to hold, no revision to pin, no way to audit what changed between one month and the next, and no path to the checkpoint if the plan lapses. The published composite score is the only durable thing you get, and it is a number about Meta's harness rather than about your workload.

The practical consequence is asymmetric and worth stating without hedging. If a request cannot leave the device — a factory floor 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 — Muse Wonder Full does not participate in the question at any price. Glimmer is a file. Full is a subscription.

What the 42-point gap is actually made of

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

• Intelligence Index — Muse Wonder Full (max) 56, fourth of 699 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. Muse Wonder Full's figure is not published on the same harness.

• Terminal-Bench 4.0 — Muse Glimmer 0.5%; the Muse Wonder series reports 85.5 and 85.6 on the vendor's table.

• Humanity's Last Exam — Muse Glimmer 22.0%; Muse Wonder Full not reported on the same harness.

• Output speed — Muse Glimmer 143.8 tokens per second. Muse Wonder Full is not published on the same harness, and neither is its cost per index task, because there is no per-token rate to compute it from.

• Cost per index task — Muse Glimmer about $0.06 hosted. Muse Wonder Full: not applicable, no metered path.

The long-context line is the one to pause on. A 30-billion-parameter model quantized into 20 GB matching a current frontier checkpoint at finding a fact inside 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 Glimmer reads well, answers fast, runs anywhere, and cannot execute. Muse Wonder Full executes and reports the highest composite score in the family — and then reports 13.1 on the legal agent suite, which is below every other frontier model in Meta's own table. Neither of those facts is explained by the other.

The $0.06 number is a trap, and so is the $50

The hosted cost-per-index-task figure deserves immediate disqualification. $0.06 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 an accelerator you already own or have to buy, plus the engineering time to serve it. Quoting $0.06 against a frontier model's per-task cost as though the two were like-for-like is the single easiest mistake to make on this page.

The $50 is a different trap in the opposite direction. It reads cheap next to a $20-per-million output rate, and it is not, because it buys no measurable unit. Ten prompts and ten million prompts cost the same, and the score you are buying it for — 59.1 on Meta's index — is not the score the open board publishes for the same checkpoint. Meta's table says 59.0 on the Artificial Analysis Index; Artificial Analysis publishes 56. A three-point disagreement on the same named index tells you which number is a claim and which is a measurement.

The honest version is that these two models are not on the same axis at all. Muse Glimmer is priced by the accelerator and its cost does not scale with volume — expensive at ten requests a day, remarkably cheap at ten million. Muse Wonder Full is priced by the month and its cost does not scale with anything you can measure. One moves with your usage. The other does not move at all.

The rung that is actually between them

Most teams reading this far do not want either extreme. They want a small model on their own hardware for the cheap, private, high-volume half of the workload and something frontier-class over an API for the hard half, and the awkward part is that those normally mean two separate stacks with two separate billing relationships.

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. That makes it the middle rung: considerably more capable than the 30B student on execution-shaped work, a fifth of Muse Wonder's output rate, and reachable on the same key as everything else.

Two honesty notes, because the naming invites the wrong assumption. Muse Glimmer itself is not one of our routes — there is no muse-glimmer entry in the catalogue, and there should not be, because Glimmer's whole point is that you run it yourself and never route it anywhere. Muse Wonder Full is not ours either, and unlike Glimmer it is not available to anyone outside Meta's subscription surfaces.

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 composes a cheap model and an expensive one into a single call, so the expensive side is invoked only where the cheap one's confidence drops. For a family with a 30B file at one end and a subscription-gated checkpoint at the other, that composition is the point — most traffic never needs either extreme.

Choosing, and the one row that decides it

• 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 or the wider series when the task has to be executed — agents, computer use, terminal work — or when the input is speech or video. A 0.5% execution score is not a budget option; it is a non-participant.

• Muse Spark 1.2 if you are shopping in the middle. It is the model Glimmer was distilled from, it is the rung between the two extremes, and it costs a fifth of what the frontier series charges for capability somewhere in the gap.

• Not Muse Wonder Full, on the strength of the composite alone. Meta reports 13.1 on the legal agent suite against 55.2 for Claude Opus 5.5 in the same table, and the independent board puts the checkpoint four points above the callable Muse Wonder rather than 42. If your work looks like that Harvey row, the highest score in the launch table is the wrong reason to pay for the subscription.

The 42-point index gap is real and it is also the least useful thing on this page. These are not adjacent products with a price step between them; they are a file and a door. The decision is where the tokens run, then whether the work is reading or executing — and the last decision, 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.