Muse Spark vs Muse Glimmer: What Actually Survives When a Frontier Model Is Distilled Into 30 Billion Parameters

The relationship between Muse Spark and Muse Glimmer is not the one the phrase "vs" implies. Muse Glimmer is not a competitor to Muse Spark — it is a student of it. Meta trained the 30-billion-parameter open-weight model by logit distillation from Muse Spark, the natively multimodal frontier line that Meta Superintelligence Labs introduced on April 8, 2026 and now ships as Muse Spark 1.3, released September 2, 2026 with its max reasoning level opened to developers two days later. Muse Glimmer followed on August 10, 2026 under Apache 2.0. So this page is not a fight card between two rivals from different labs. It is a measurement of how much of a frontier model survives a sixteen-fold shrink and a licence change — and Meta, unusually, has published numbers on both sides of that shrink. Whether that comparison is fair is the first thing worth being honest about, and the answer is: partly.

One is a file, the other is an endpoint

Start with the part that has nothing to do with benchmarks, because it decides most real deployments before a single score matters.

Muse Glimmer is a download. Meta publishes the weights under Apache 2.0, ships them already quantized to 4-bit — which takes the footprint from over 55 GB at full precision to under 20 GB — and states that the model is small enough to run on a single consumer GPU or a Mac. It runs with the network unplugged. Meta tested it on an Nvidia RTX 5090 and on Apple M4 Max and M5 Max silicon. It accepts text and image input, returns text, carries a 131,072-token context window that Meta says can be extended to 262,144, and has a knowledge cutoff of January 2026.

Muse Spark is the opposite shape at every point. Meta publishes no weights, no licence and no download for any version of it; it is reachable only over Meta's Model API, at meta.ai and in the Meta AI app. It runs a 1,048,576-token context window — eight times Glimmer's, before Glimmer's extension. It accepts text, image, video, audio and PDF input and returns text. It is priced per token: $1.25 per million input, $0.15 per million cached input, $4.25 per million output, with a Contributor tier at $0.10 / $0.002 / $0.20 that trades a lower price for permission to train on your prompts and completions. Glimmer has no per-token vendor bill at all when you run it yourself, because there is no vendor in the loop.

• Size — Muse Spark: undisclosed, Meta publishes no parameter count for any version; Muse Glimmer: 30B, shipped at 4-bit under 20 GB

• Weights — Muse Spark: proprietary, hosted only; Muse Glimmer: Apache 2.0, downloadable and runnable offline

• Context window — 1,048,576 tokens vs 131,072, extendable to 262,144

• Inputs — text, image, video, audio, PDF vs text and image

• Output — text vs text

• Price — $1.25 / $0.15 cached / $4.25 per million tokens vs no vendor charge when self-hosted

• Knowledge cutoff — not published for Muse Spark 1.3; January 2026 for Muse Glimmer

The one benchmark Meta published for both

This is the rarest kind of number to find in a distillation relationship: the same evaluation, run by the same lab, reported on the teacher and the student.

Meta's model pages list GDPval-AA at 1,754 for Muse Spark 1.3 running its max reasoning configuration, and 953 for Muse Glimmer 30B at high reasoning. That is the student retaining 54% of the teacher's score on that task family — a real, published datapoint rather than an inference. Two caveats belong on it immediately, because both are true and neither is small. First, the Muse Spark row is labelled v2 and the Muse Glimmer row is not versioned at all, so if Meta has moved the benchmark between the two measurements the ratio is soft even though the direction is not. Second, both figures are vendor-reported. We have not reproduced either, and no independent harness has published a head-to-head on this pair that we could find.

What makes the 801-point gap legible rather than alarming is the third-party picture underneath it. On Artificial Analysis' Intelligence Index, Muse Spark 1.3 at max scores 48 and Muse Glimmer at high scores 17. Those are different index scales from different measurement rounds and the two are not ranked against the same pool — Artificial Analysis buckets open-weight models by size class, so Glimmer is measured against open-weight peers while Muse Spark is measured against the whole field. Read them as shape, not as a ratio. The shape says the same thing Meta's own table says: the distilled model is a genuinely different capability tier, not a cheaper route to the same answers.

What the small model keeps

The interesting part of distillation is what survives the compression, and on Meta's own published rows Muse Glimmer is not uniformly beaten. It posts 76.0 on SWE-Bench Verified, 65.9 on OSWorld-Verified, 77.0 on IFBench, 94.7 on AIME 2026 and 83.5 on GPQA Diamond. Those are strong numbers for a 30B model and they are the reason the architecture choice matters: Meta describes Glimmer as fine-tuned on long-context agentic data and reinforced specifically for tool use, with self-managed memory across hours-long sessions and recovery from a failed tool call rather than a halt. That is a targeted model wearing a general model's name.

The most useful piece of context on the page is Meta's own choice of comparison. Its benchmark table for Muse Glimmer does not put it next to Muse Spark. It puts it next to Gemma 4 31B in thinking mode and Qwen3.6 27B in thinking mode — same size class, same open-weight premise. On that table Glimmer wins most rows but not all: it takes MCP Atlas at 75.5 against 62.5 for the Qwen model, TerminalBench 2.1 at 51.7 against 60.7, and GDPval-AA at 953 against 1,141. So even inside its own weight class the distilled Meta model is not the automatic pick, which is a more honest framing than the page's headline suggests.

Two further rows are worth pulling out because they are measurable and easy to get wrong in planning. Artificial Analysis records Muse Glimmer's cost per completed Intelligence Index task at $0.05 and its output speed at 122.4 tokens per second — against $1.60 and 121.9 tokens per second for Muse Spark 1.3 at max. On that pair of figures the small model is roughly 32 times cheaper per completed task at essentially identical throughput. That is the distillation paying off, and it is invisible if you only compare index scores.

The place the comparison breaks

It breaks on inputs, and it breaks harder than the scoreboard suggests.

Muse Spark 1.3 takes video and audio and PDFs. Muse Glimmer takes text and images. If your pipeline feeds a screen recording, a call recording or a scanned document straight into the model, the two are not alternatives at any price — one of them physically cannot accept the payload. Meta's own documentation reinforces the point from an unexpected direction: it warns that audio understanding in Muse Spark 1.3 is not fully supported and that quality on audio-carrying requests may be degraded, recommending Muse Spark 1.2 for audio workloads. So even within the hosted line the newest version is not the right answer for every input type, and the audio-capable option is the older checkpoint.

The second break is context. One million tokens versus 131,072 is not a marketing difference; it is the difference between a model that can hold a large repository or a long case file in a single call and one that needs a retrieval layer and a summarisation step built around it. If your work is long-horizon codebase reasoning, Glimmer's context ceiling is the constraint you will hit first, long before its benchmark scores.

The third is data boundary, and it cuts the other way. A request that never leaves your hardware cannot be logged by a vendor, cannot be priced per token, and cannot be affected by a rate change. For a workload with residency requirements or a per-token budget that does not survive volume, that single property outweighs 30 index points.

Where each one actually lives

Muse Glimmer's distribution is the wide one, because open weights are. Meta lists the download on Hugging Face and points at local runtimes — llama.cpp, MLX, Ollama, LM Studio, Unsloth — alongside the vendor's own hosted list and several third-party inference platforms. We will not name those platforms here; the point for a reader is simply that the weights are portable, and anything that can load a 30B 4-bit model can serve it.

The Muse Spark line is narrow by construction. Three versions, one vendor, one API, no weights. That concentration is the real risk of building on it, and it is a risk that does not apply to Glimmer at all.

Both on one key, where that is possible

Our own catalogue is narrower than either of these stories. We route Muse Spark 1.1 and Muse Spark 1.2 — the two checkpoints listed at $1.25 in and $4.25 out per million tokens, both at Meta's list price with our 0% markup passed through untouched — and we do not host Muse Spark 1.3, either Contributor variant, or Muse Spark Open. Muse Glimmer is not on our catalogue either, and the honest advice for that one is different in kind: it is open weights, so the natural place to run it is your own hardware or a runtime that loads it.

Where routing does earn its place here is the migration you are probably considering. If you are testing whether a 30B distilled model can take over part of a frontier workload, the thing you want is both in one endpoint so a slice of traffic can be moved and measured without a second contract — Muse Spark 1.1 and 1.2 on one key, with automatic failover so a failed long-context call degrades to the other checkpoint instead of to an error, and a routing DSL for splitting traffic by configuration rather than by code. When Meta's hosted line moves again, the change lands in one place.

Which one to pick

If your work is text and images, runs on your own machine or a rented GPU, and needs to be immune to per-token pricing, Muse Glimmer is the answer and it is not close — you are trading roughly half the GDPval-AA score for a model that costs nothing per call and cannot be rate-limited by anyone.

If your work involves video, audio, PDFs or a genuinely long context, Muse Glimmer is not a cheaper version of Muse Spark. It is a different tool, and the distillation did not carry those capabilities across. Reach for the hosted line, budget $1.25 and $4.25 per million tokens, and check which checkpoint actually handles your payload — for audio, Meta's own documentation points at the older version.

The honest summary of a distillation comparison is that the student is remarkably good at the thing it was trained to do and simply absent from the things it was not. Muse Glimmer is the strongest argument for open weights Meta has published since Llama 4 and the numbers above are Meta's own. What they do not show — and no vendor table can — is how the pair behaves on your workload. That is an experiment, not a lookup.

Benchmarks age and prices move; the failure record does not: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.