Muse Spark Open vs Muse Spark: What You Give Up When Meta Hands You the Weights

The comparison people actually mean when they type "Muse Spark Open vs Muse Spark" is not two models — it is two distribution models wearing the same family name. Muse Spark Open is the open-weight 816B-parameter sparse mixture-of-experts model Meta published on October 15, 2026, downloadable under the Muse Model License, running text, image, video and audio in and text out. Muse Spark is the closed hosted line of the same family — the original model from April 2026, then Muse Spark 1.1, 1.2 and 1.3 in sequence — sold per token through Meta's Model API and reachable no other way. One arrives as a file and a licence; the other arrives as an endpoint and a rate card. Almost every practical difference between them follows from that single fork.

First, a naming problem worth clearing up

"Muse Spark" without a version number is genuinely ambiguous, and the ambiguity gets worse now that an open sibling exists.

• Muse Spark, unversioned — the original model, released April 2026, 262K-token context on Artificial Analysis' listing, text, image and speech in.

• Muse Spark 1.1 and 1.2 — July and August 2026 respectively, both at the family's 1M-token context window.

• Muse Spark 1.3 — September 2026, the version Meta labels current and recommends for new work.

• Muse Spark Open — a separate open-weight release, not a version of the closed line, and not a successor to 1.3.

And a fourth thing that is often confused with all of the above: Muse Glimmer, a 30B dense open model under Apache 2.0 from the same research group. Someone searching for "open Meta Muse model" and landing on Glimmer has found a real open-weight model and the wrong one. Muse Spark Open is the 816B model; Glimmer is the small local one.

What the two sides are, on paper

Six dimensions, both sides, using the vendor's stated figures and Artificial Analysis' independent listings where one exists.

• Weights — Muse Spark Open: downloadable, 816B total / 42B active, gated repository, BF16 and NVFP4 checkpoints. Muse Spark (closed line): no weights, no licence, no download for any version, including 1.3.

• Inputs — Open: text, image, video and audio per Meta, and text, image, speech and video per Artificial Analysis' listing. Closed 1.2: text, image, speech and video per Artificial Analysis. The two sources agree on the open model's modality list; what neither has published is a measurement of the audio or video path.

• Output — text only, both sides. Neither generates images, audio or video.

• Context — 1,048,576 tokens for Open and for 1.1, 1.2 and 1.3. The original unversioned Muse Spark sits lower, at 262K.

• Licence and terms — Open: Muse Model License with an attribution requirement and a model-as-a-service clause that applies above $20M in annual affiliate revenue. Closed: commercial API terms, with a discounted "Contributor" tier that trades a lower price for permission to train on your prompts and completions.

• Price — Open: Meta sets none, because Meta does not serve it. What you pay is either your own hardware or whatever an inference platform charges. Closed 1.3: $1.25 per million input tokens, $4.25 per million output, $0.15 per million cached input, with a Contributor variant at $0.10 / $0.20 / $0.002.

The independent numbers, and the honest caveat about them

Artificial Analysis is the only third party with public listings for both sides of this comparison, which makes it the only place a like-for-like read is even possible — with one asterisk it does not hide: it benchmarks open-weight models against other open-weight models of the same size class, and proprietary models against proprietary and open models in the same price band. Note also that the closed line's figures below are Artificial Analysis' own measurements of hosted endpoints, while the open model's Index and speed come from the same analyst's hosted listing of a model nobody is obliged to host. The scores are therefore comparable in direction, not audited equivalents.

• Muse Spark Open (xhigh) — Intelligence Index 39, text, image, speech and video in, output speed 103.4 tokens per second, listed hosted pricing $1.25 in / $4.25 out, $0.84 per index task.

• Muse Spark 1.2 (xhigh) — Intelligence Index 40, text, image, speech and video in, 361.4 tokens per second, $1.25 in / $4.25 out, $0.97 per index task. Its page now carries a deprecation notice pointing at 1.3.

• Muse Spark 1.3 (max) — Intelligence Index 48, text, image and video in, 121.9 tokens per second, $1.25 in / $4.25 out, $1.60 per index task.

• Original Muse Spark — Intelligence Index 31, 262K context, speed and cost unmeasured, marked deprecated.

The shape of that list is the interesting part. The open model does not beat the closed line. It lands one point below the deprecated 1.2 and nine points below 1.3 at its maximum reasoning setting — and it does that while being roughly three and a half times slower than 1.2. What it does win is cost per task: $0.84 against 1.3's $1.60 on the same index, which is the cheapest per-task figure in the family. So the trade is not capability and it is not really price-per-token, since every hosted listing in this table quotes the same $1.25 / $4.25. It is control.

One number in that table deserves scepticism rather than quotation. Muse Spark Open's listed hosted price is identical to Meta's list price for the closed Muse Spark 1.3, to the cent, on all three columns. That could be a genuine market clearing price, or it could be an analyst filling the field with the sibling's rate card while the model's own pricing settles. Until an actual provider invoice confirms it, treat Open's cost figures as provisional and 1.3's as vendor-published.

What the open licence costs you that the API never would

The licence is where the two paths diverge most sharply, and it is the section most readers skim.

What you get plainly: freedom to use, copy, modify, fine-tune, distill, merge, publish, sublicense and sell derivatives, with Meta claiming no ownership of outputs. Two conditions land on top of that.

• Attribution — the licence requires "Built with Muse Spark Open" to be displayed prominently in the interface of any product using the model or a derivative, or in its documentation, site or settings menu.

• The model-as-a-service trigger — if you or your affiliates offer third parties access to the model as a product or service over an API, and aggregate gross revenue exceeds $20M across any consecutive twelve months, you must notify Meta and sign a separate agreement first. Internal use, research, and shipping the model inside your own non-API product are carved out.

• Gated access — the download itself is access-controlled, with a request form that asks for full legal name, date of birth and organisation with corporate identifiers, and cannot be edited once submitted.

Set that against the closed line's terms and the comparison inverts. The API route asks nothing about attribution, imposes no revenue threshold, and involves no paperwork beyond a key. What it does ask is the data question: on the Contributor tier, Meta may train on your prompts and completions, which is precisely what makes it roughly ten times cheaper on input and better than twenty times cheaper on output. The closed Standard tier's promise is the opposite — your data is not used for training.

What each path actually costs to run

The headline comparison of "$0 per token versus $1.25 per million" is wrong in both directions, so it is worth pricing honestly.

• Datacentre BF16 — Meta's own recipe runs the full checkpoint across four nodes of eight H200 GPUs at tensor parallelism 32. Open weights did not remove the hardware bill; they moved it onto your balance sheet.

• Datacentre NVFP4 — the 4-bit build is documented on a single B200 node, with Meta's SGLang reference configuration spanning four GB300 GPUs.

• Desktop — the genuinely new option. Two Mac Studio machines with M5 Ultra chips hold a full BF16 checkpoint, because each exposes up to 512GB of unified memory at 1.2TB/s of bandwidth. A 2-bit quantisation tested by the Hugging Face and llama.cpp teams ran on a single 512GB M3 Ultra at 18 tokens per second using 304GB, with room left for cache.

• The context trap, on both paths — at the full 1M-token window in BF16 the KV cache alone runs to roughly 266GB. Buying the weights does not buy you the window; on a hosted endpoint the same ceiling sits behind a per-token price that grows the same way.

The closed line has no equivalent to the desktop row. There is no configuration in which Muse Spark 1.3 runs on hardware you own, at any quantisation, at any speed.

Which one to pick, by situation

The decision rule is narrower than the marketing on either side suggests.

• If your workload is capability-critical and you need the best answer the family can produce today, the closed line wins and it is not close — 1.3 at maximum reasoning scores nine index points above the open model, and the open model is slower.

• If your constraint is that prompts must not leave infrastructure you control, or that you need to fine-tune, merge or distill, the open weights are the only option in the family. Neither is true of any Muse Spark version.

• If you want the cheapest per-task outcome on this family, the open model's $0.84 index task is the lowest figure on the table — but it is also the least verified number on the table, and it depends on which platform you run it on.

• If you want to evaluate either of them on real traffic before committing, the split itself is the argument for a single endpoint: Muse Spark 1.2 is one of the two Muse models OrcaRouter routes, at Meta's list price with 0% markup, so a published price change reaches your key without a renegotiation. Muse Spark Open is not on our catalogue and we do not claim it is — it is a model you download, not one you call. Nor do we route Muse Spark 1.3; the two versions we serve are 1.2 and 1.1.

What the routing layer buys in this specific case is not a discount, it is an exit. One API across 200+ models, automatic failover when a provider degrades, and a routing DSL that moves traffic between models by configuration all mean the choice between open and closed stops being a migration. A model released a day ago with one analyst index and no reproduction is exactly the case where you want the option to shift traffic without a code change behind it.

The part neither side has settled

Two open questions sit over this comparison, and neither resolves in favour of a clean verdict.

The first is measurement. Meta's own scorecard for Muse Spark Open is vendor-reported and harness-relative, and the safety report its launch post points to is described as coming soon rather than published. The independent picture is one analyst index, one speed figure and one modality listing — and that listing agrees with the vendor on the specification while saying nothing about which of those four input paths has actually been measured. Nine index points is a large gap, and it is currently a gap measured in different model classes by a single party.

The second is whether open weights change the family's direction at all. Meta has now published a model at this scale while keeping every version of the closed line proprietary, and the open release sits below the closed flagship on capability rather than level with it. That is a coherent strategy — the frontier stays hosted, the previous generation's shape becomes a download — and it is one Meta has not stated. If a future Muse Spark Open matches the closed line at release, this comparison stops existing. Until then, the honest reading is that open weights bought you the family's architecture and its context window at a discount on capability, and it did not buy you the frontier.

Comparisons like this one are also where the failure modes show up: 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.