Muse Spark Open vs Muse Glimmer: What 22 Index Points Actually Cost You

Thirty-nine against seventeen. That is the whole of the comparison between Muse Spark Open and Muse Glimmer on the Artificial Analysis Intelligence Index, and it is the number that travelled after Meta published the larger model's weights on October 15, 2026. On its own it is close to useless, because the two models are not two answers to the same question. One is an 816-billion-parameter mixture of experts built to be served from a data centre. The other is a 30-billion-parameter dense model built to run on a single consumer graphics card. Reading the gap as "the new one is better" gets the direction of the decision right and the reason for it completely wrong. None of this is a reason to avoid the model, only a reason to know the history: 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.

Here is what the gap actually buys, what it costs, and the one axis on which the smaller model still wins outright.

Two downloads that answer different questions

Both carry Meta's name, both put weights on the Hugging Face Hub, and both let you run inference without sending a token to anyone's endpoint. That is where the resemblance stops.

• Muse Spark Open — 816B total parameters, 42B active per token, sparse mixture of experts, published October 15, 2026 under the Muse Model License.

• Muse Glimmer — roughly 29.6B total parameters, dense, plus a separate 1.8B vision encoder, published in August 2026 under Apache 2.0.

The active-parameter difference is smaller than the headline suggests: 42B against roughly 29.6B is a factor of about 1.4, because Muse Spark Open only wakes a slice of its experts per token. The total-parameter difference is what funds the hardware bill: about 27.6 times the weights to store, quantise and move.

What 22 points costs per task

Artificial Analysis prices the two models against the same yardstick — the cost of completing one task on its own Intelligence Index — and that is where the gap gets a second dimension. On its hosted listing, Muse Spark Open at xhigh reasoning runs $0.84 per Index task, against $0.05 for Muse Glimmer at high. That is roughly 16.8 times the cost for 2.3 times the score.

List pricing tells the same story in the units a finance team uses. Muse Spark Open is listed at $1.25 per million input tokens and $4.25 per million output tokens, with an 88% cache discount and a blended rate of about $0.78 per million tokens. Muse Glimmer is listed at $0.30 per million input and $1.20 per million output. If your workload is a high-volume classification, extraction or routing pass, the smaller model is not a compromise — it is the correct answer priced at a fifth of the input rate.

Output speed is the one metric where they are close enough to be a tie, and it is worth stating plainly because it is counter-intuitive: Muse Spark Open measured 102.0 output tokens per second with a 0.67-second time to first chunk, Muse Glimmer 124.4 tokens per second at 1.14 seconds. The 816B model is not the slower one to answer a single prompt, because it activates a fraction of itself and runs on hardware sized for it. Where the smaller model wins is at the edge: a 4-bit build drops the language model under 20 GB, and a companion block-diffusion drafter that proposes 16 tokens at a time is claimed to give up to 3.1× speedup on an RTX 5090. That is a vendor figure and has not been independently reproduced.

The context window is not a fair fight either

Muse Spark Open carries a 1,048,576-token window and accepts text, images, video and audio in, with text out. Muse Glimmer carries 131,072 tokens — 128k in the round numbers Meta writes — and takes text and images, with no audio path at all. The perception encoder in Muse Spark Open is a 1.8B, 50-layer component with sparse attention, and the language stack is 62 layers, hidden dimension 8,192, a 202,048-token vocabulary, 256 experts with top-8 routing and grouped-query attention at 128 query heads against 16 key-value heads.

If your problem is "read this two-hour recording and this stack of screenshots and tell me what happened," Muse Spark Open is the only one of the two that can hold the input. If your problem is "read this one screen and click the right button on my laptop," 128k is not the constraint you will hit first.

The licences are not the same kind of open

This is the part the 22-point headline buries, and it is the part a legal review will care about.

• Muse Glimmer — Apache 2.0 across everything Meta shipped: full-precision BF16 weights, two 4-bit quantised variants, the drafter head and the perception encoder. Commercial and research use, no gate.

• Muse Spark Open — the Muse Model License. Broad rights to use, copy, modify, fine-tune, distill, merge and sell, subject to three conditions worth knowing before you plan: copies and derivatives have to carry the licence and usage policy; products built on it have to display "Built with Muse Spark Open" in the interface, documents or settings; and if you offer third-party access to models whose outputs those third parties receive, and combined gross revenue exceeds $20 million over any consecutive twelve months, you must notify Meta and sign a separate agreement first. Internal use, research and non-model-as-a-service commercial work are carved out. Meta claims no ownership of outputs.

Access is also different in kind. The Muse Spark Open repository is gated — you log in and submit your full legal name, date of birth and organisation name with identifiers, and the page warns the form cannot be edited afterwards. Muse Glimmer downloads are not gated. For a team that wants to evaluate both quietly before committing, that asymmetry is real friction, and the $20M clause means "we downloaded it" and "we can ship a hosted product on it" are two different answers.

Where each one fits on the rack

Meta's own model card states the hardware plainly: the BF16 checkpoint needs about 2 TB of VRAM, and the NVFP4 build needs at least 700 GB. Supported serving stacks are SGLang, vLLM, Hugging Face, TensorRT-LLM, Unsloth and llama.cpp, and the card also points at MLX tuned for two M5 Ultra Mac Studios. That is the desktop demo Meta led with — two top-spec machines chained together, not a laptop.

Muse Glimmer's target is genuinely different: a single RTX 5090, an under-20 GB quantised build, and a drafter head designed to make the tokens come out faster on that card. The two models do not compete for the same silicon. Meta's launch post for the larger model makes the same point from the other direction, leading with agentic work and the two-Mac-Studio configuration rather than with a single-card number.

Which one to pick

The decision collapses quickly once you fix the constraint.

• You are shipping a local agent on hardware you own — Muse Glimmer. The 22 points are unavailable at any price you can pay on one GPU.

• You need audio, video or a million-token input in one call — Muse Spark Open, and no smaller model substitutes.

• You are doing high-volume, low-difficulty work — Muse Glimmer at $0.30 per million input tokens, unless a measured evaluation says the 17-scoring model fails your specific task.

• You want the frontier score and you are renting — the honest note is that Muse Spark Open is not the cheapest way to reach a 39. It is the only way to reach a 39 with the weights on your own disk.

On the serving question, the practical reality for most teams is that neither model is one you self-host on day one. If you want to call Meta's Muse line without standing up GPUs, Muse Spark 1.2, which is one of the two Muse models OrcaRouter actually routes, is available through one API for more than 200 models, with automatic failover so an unfamiliar model can carry test traffic without carrying your production path. Both Muse models we serve are billed at provider list price at 0% markup, so a price change on the vendor side is live with us the same day. To be explicit about the limits: OrcaRouter does not currently route Muse Spark Open, and does not route Muse Glimmer — both are weights you run yourself, or endpoints you buy elsewhere.

What is still unmeasured

Almost everything that would settle the comparison.

The 39 is one independent number attached to a model whose own launch scorecard lists JobBench 61.8, MCP-Atlas 89.2, GDPval-AA v2 1401, OSWorld Verified 80.4, Toolathlon Verified 72.8, HLE-Diamond 26.1 without tools and 61.0 with them, GPQA Diamond 89.0, SciCode 58.5, DeepSWE 1.1 at 60 and Terminal-Bench 2.1 at 76.4 with the Terminus 2 agent. Every one of those is vendor-reported and unreproduced, and several were measured inside Meta's own agent harness, which makes them a capability claim rather than a like-for-like ranking. Glimmer's card carries its own vendor table — SWE-Bench Verified 76.0, Terminal-Bench 2.1 51.7, GPQA Diamond 83.5, AIME 2026 94.7, HLE 22 — with the same caveat. Two vendor scorecards produced months apart on different harnesses cannot be subtracted from each other, which is exactly why the single independent Index is the number this piece leans on.

The gap that will actually move the decision is not on either card. It is the cost of a task at the quality bar you personally need, on your own inputs, measured once. Until somebody runs that, 39 and 17 are two facts about two different machines wearing the same family name.