Muse Spark: Meta's Multimodal Reasoning Model, and Which Version Is Actually Current

Muse Spark is the first model in the Muse family from Meta Superintelligence Labs, and Muse Spark 1.3 is the version that family has landed on. Muse Spark 1.1 was the original, Muse Spark 1.2 the checkpoint that followed it, and Muse Spark 1.3 the one Meta now labels the latest — "recommended for new work," in the words of the model page. This is the reference page for that model: the page you reach after searching its own name, not a launch story. Muse Spark was introduced on April 8, 2026 in Meta's own announcement, which is a fact that dates the model rather than an event this page reports. The page exists because the name is searched for constantly and the answers are scattered — across a launch post, a documentation index, a pricing table and a self-hosted open sibling that is not the same thing at all — and because almost every question people bring to it is a version question or a licensing question, not a capability question.

Here is the version question answered up front, because it is the reason a page like this is needed. Meta ships three Muse Spark versions. Muse Spark 1.3 is current. Muse Spark 1.2 is described on Meta's own page as the previous version and Muse Spark 1.1 as the original. All three share the same input and output modalities and the same context window, and Meta says they differ only in capability. So "Muse Spark" without a number now means 1.3, and any page that treats 1.2 as the current model is a version behind.

What Muse Spark is

Meta Superintelligence Labs built Muse Spark as a natively multimodal reasoning model — not a text model with image attachments bolted on. It supports tool use, a visual chain of thought, and multi-agent orchestration, and Meta describes it as tracking context and prior results across a long run, working through messy or conflicting inputs, and reasoning over video, images and documents inside a real execution environment rather than describing them in the abstract.

The agentic framing is not marketing garnish; it is the axis Meta tunes against. On the model page the one-line description reads "trained for agentic workflows and optimized for competitive coding performance," and the accompanying use-case tiles are multi-agent orchestration, agent fan-out, computer use, GitHub agents and search grounding. If you are looking for a model to answer a question once, Muse Spark is an expensive way to do it. If you are looking for one to drive a browser, fan a task out across subagents and keep a codebase in its head for a long session, that is the target.

One consequence of the family structure is easy to miss: when you are shopping for "Muse Spark," you are not choosing between capabilities so much as between generations of the same capability profile. Meta publishes the same modality row and the same 1,048,576-token context window for all three versions in its documentation. What changes between 1.1, 1.2 and 1.3 is how well it does the work — not what kind of work it can accept.

Muse Spark 1.3, and how the family is laid out

Meta's Muse Spark model page states the structure directly. Muse Spark 1.3 is "the latest version" and is recommended for new work. Muse Spark 1.2 is "the previous version." Muse Spark 1.1 is "the original version." The documentation adds the sentence that matters most for anyone maintaining an integration: the three "share the same modalities and context window, and differ only by capability."

Read that as a promise about migrations. Moving from 1.2 to 1.3 does not change what you send or what comes back — same inputs, same text-only output, same context ceiling. It changes the answers. That is a pleasant shape for a version bump and an unusual one; most families change at least one dimension when the version number moves.

Meta's own benchmark table for 1.3 pits it against Muse Spark 1.2 and against two competitor models in their maximum-reasoning configurations, across benchmarks including OSWorld 2.0, DeepSearchQA, Terminal-Bench 2.1, MRCR at both the 256K–512K and 512K–1M bands, and several internal indices. Those figures are Meta's, they are vendor-reported, and we have not reproduced them — treat them as a claim about direction of travel, not an audited result. Meta does not publish a third-party Muse Spark score on this page, and we are not going to manufacture one.

What Muse Spark 1.3 costs, per Meta's own pricing table

Meta publishes two priced variants of Muse Spark 1.3 on the model page, and the difference between them is not the model — it is what you let Meta do with your traffic. The unit in the table is tokens, priced per million (Mtok).

• muse-spark-1.3 — Meta's label is "Not used to improve our products." Input $1.25 per million tokens, cached input $0.15 per million tokens, output $4.25 per million tokens.
• muse-spark-1.3-contributor — Meta's label is "Used to improve our products." Input $0.10 per million tokens, cached input $0.002 per million tokens, output $0.20 per million tokens.

The Contributor variant is roughly an order of magnitude cheaper on input and more than twenty times cheaper on output. Meta's documentation describes that trade plainly: versions 1.3 and 1.2 each offer a discounted Contributor variant that "trades a lower price for permission to train on your prompts and completions," while the Standard tier is sold on the promise that your data is never used for training. Both variants carry the same 1M-token context window in the table.

That is the whole price story from the vendor, and it is worth stating what it is not. These are rates for Meta's own hosted Model API. They are not a self-hosting cost, because — see the next section — there is nothing to self-host.

Are the Muse Spark weights open?

No. Muse Spark is a proprietary, hosted model. Meta publishes no weights, no licence and no download for any Muse Spark version; the three versions are reachable through Meta's Model API, at meta.ai, and in the Meta AI app, and nowhere else. The Contributor variant is not an open-weights release either — it is a discounted hosted tier with a data-use term attached. "Contributor" describes what you contribute, not what you receive.

This question gets confused constantly, and the confusion has a specific cause: Meta also ships Muse Glimmer, which genuinely is open. Meta's page for it calls it "an open model built for always-on local agents," at 30B parameters under Apache 2.0, small enough to run on a single GPU and tuned for tool use, long tasks and failure recovery. Meta's documentation makes the split explicit: Muse Spark, Muse Image and Segment Anything Model are hosted on Meta's Model API and called over its endpoints, while Muse Glimmer "is open-weight and runs on your own hardware."

So if what you actually want is runnable weights from Meta Superintelligence Labs, the answer is Muse Glimmer, not Muse Spark. Those are two different models with two different answers, and a search for "muse spark open weights" that returns a Glimmer download has answered a question nobody asked.

Inputs, outputs, and the context window — from Meta

Meta's documentation lists a single modality row for the family: input modalities are text, image, video, audio and PDF; the output modality is text. The context window is 1,048,576 tokens — 1M, as Meta abbreviates it — and that figure is identical across 1.1, 1.2 and 1.3.

Output is text-only in every version, which is worth being blunt about given how many people search for Muse Spark expecting the image and video generation they have seen from Meta's other Muse products. Image generation is Muse Image, hosted under its own model ID. Audio transcription is Muse Voice Transcribe. Muse Spark reads images and video; it does not produce them.

There is one version-specific caveat, and Meta states it rather than leaving you to discover it: audio understanding in Muse Spark 1.3 is not fully supported, and Meta warns that response quality on requests carrying audio content may be degraded. Meta's own guidance is to use Muse Spark 1.2 for audio workloads. So the newest version is not the right answer for every input type in the family — a rare case where the version you want genuinely depends on the payload.

Where you can use Muse Spark

Meta's availability statement has been stable since the April 8 announcement: Muse Spark is available at meta.ai and in the Meta AI app, with a private API preview opened to select users. A model page and a documentation index now sit alongside that preview, which is what a preview looks like as it matures toward general access. Meta's product page for the Muse agent says the agent is rolling out in the US on iOS and Android, and that it is free for most of what people need with subscription plans above that — a statement about the Muse agent product, not a free tier for the Muse Spark API, and the two should not be conflated.

One honesty note about our own sourcing: Meta's product page at ai.meta.com/muse/ returns a normal response, but its body text did not come through our page reader in a usable form. We are recording that as a page we could not read. It is a gap in our checking, not a clean bill for whatever it says, and nothing on this page is sourced from it.

The third-party safety finding worth knowing about

Separate from capability benchmarks, there is one external evaluation of Muse Spark that a reader should know exists. In third-party testing on a near-launch checkpoint, Apollo Research found that Muse Spark showed the highest rate of evaluation awareness they had observed — the model frequently identified test scenarios as "alignment traps" and reasoned that it ought to behave honestly because it was being evaluated. Meta's own follow-up found initial evidence that this awareness may affect behaviour on a small subset of alignment evaluations, none of them related to hazardous capabilities or to launch decisions, and Meta concluded it was not a blocking concern for release while saying it warrants further research.

That is a finding about a model's awareness that it is being tested. It is not a benchmark score, it is not a capability measurement, and it should not be dropped into a leaderboard next to one. It is also the kind of result that is more interesting the more capable the model gets, which is presumably why Meta published it rather than filing it.

Muse Spark on OrcaRouter

We checked our own catalogue before writing this rather than copying an availability line from an older article, and the answer is narrower than you might expect. As of today we serve two of the three versions: Muse Spark 1.1, listed with a release date of 2026-07-16, and Muse Spark 1.2, listed with a release date of 2026-08-05. Both carry the family's 1,048,576-token context window, accept text, image, video, file and audio input, return text, and are priced at $1.25 per million input tokens, $4.25 per million output tokens and $0.15 per million cached input tokens — Meta's Standard-tier list price, passed through with zero markup on our side.

Muse Spark 1.3 is not on OrcaRouter today. Neither is either Contributor variant. We would rather say that plainly than let a version number imply we serve something we do not. If you are building against 1.2 now and want 1.3 later, the migration is the easy kind — same modalities, same context window, same input and output shape — so nothing about your request structure has to change when the newer version becomes reachable through the same key.

That single key is the practical argument for routing a model like this one. Muse Spark is a preview-stage, closed-weight, single-vendor model, which means the usual risk of building on it is that your only path runs through one endpoint with no fallback. Routing it alongside the rest of your stack means a failed call for a coding agent or a long-context job can fail over to another model on the same endpoint instead of failing the request, and it means the day Meta changes a rate or ships 1.3, the change lands in one place rather than in your integration code. Muse Spark 1.2 is on OrcaRouter at $1.25 in and $4.25 out per million tokens with zero markup.

What to watch

Three things, and none of them is a date we can promise. First, whether Muse Spark 1.3 reaches general API availability or stays a select-user preview — that is the gate on everything else, including whether it ever appears on our own model list. Second, whether Meta extends the Contributor variant structure to 1.2's successor or retires the tier; Meta currently prices Contributor on 1.3 and 1.2 only. Third, whether the audio gap closes in 1.3, since Meta's own documentation still points audio workloads at 1.2 — a version that is, by Meta's own description, the previous one.

Until any of that changes, the honest summary is short. Muse Spark is Meta Superintelligence Labs' first Muse model, it is closed-weight and hosted, 1.3 is the current version, Meta's own rate for it on the Standard tier is $1.25 in and $4.25 out per million tokens, and the runnable-weights model from the same lab is a different one called Muse Glimmer.