What Is Muse Spark?

Muse Spark is Meta's family of multimodal reasoning models, built by Meta Superintelligence Labs. It's the model behind Meta AI, the Meta Muse agent, and the Muse Code coding agent, and developers can call it through Meta Model API. Meta introduced it on April 8, 2026. The current version is Muse Spark 1.3, released September 2, 2026.

• Versions you can call: muse-spark-1.3 (latest), muse-spark-1.2, and muse-spark-1.1. Versions 1.3 and 1.2 also come in cheaper Contributor variants.
• Context and modalities: 1,048,576 tokens on every version. Input can be text, images, video, audio, and PDFs. Output is text only. Meta says audio isn't fully supported in 1.3 yet.
• Price per 1M tokens: $1.25 input, $0.15 cached input, and $4.25 output on Standard. Contributor costs $0.10, $0.002, and $0.20, and in exchange Meta can train on your prompts and completions.
• Where to use it: Meta Model API, Muse Code, OpenRouter, and Meta's consumer products (meta.ai, the Meta AI app, and Meta Muse).
• Open weights: No. You can't download or self-host Muse Spark today. Meta lists "the Muse Spark open weights release" on its roadmap without a date. Muse Glimmer, a smaller model distilled from Muse Spark, is Meta's open-weight model (Apache 2.0).
• How good it is: Meta's own charts put 1.3 at or near the top on several coding and long-context benchmarks. Those are vendor numbers, and Meta's pages disagree with each other on some scores. Outside evaluations are more mixed (see below).

Muse Spark versions and release timeline

All three API versions are still served. Meta's Models docs recommend 1.3 for new work and call 1.2 "the previous version." They call 1.1 "the original version," even though the April model came first. The April release had no API model ID, which may be why Wikipedia gives July 9 (the 1.1 date) as the release date.

Meta hasn't announced any version after 1.3 on the pages we read.

Models, context window, and modalities

* Meta's Models docs say audio understanding in 1.3 "is currently not fully supported" and that answers may be worse when a request includes audio. For audio, Meta points you to Muse Spark 1.2, or to Muse Voice Transcribe if you only need speech-to-text.

Some older third-party pages list smaller context windows. Meta's current docs give 1,048,576 tokens for every served version, and Meta doesn't charge more for long context.

Muse Spark doesn't generate images or speech. Meta sells those as separate models (Muse Image and Muse Voice Transcribe).

Reasoning levels

Muse Spark always reasons before it answers. Meta Model API returns an error if you try to turn reasoning off with none. You pick the depth with reasoning_effort, from minimal through low, medium, high, and xhigh up to max.

On Meta Model API, max is limited to Standard-tier muse-spark-1.3. Contributor models can't use it. Muse Code's configuration docs say Muse Code offers max on both Standard and Contributor. That's how Meta documents each surface. We haven't tested which one is right.

Pricing and the Contributor tradeoff

Source: Meta Model API pricing docs and the Muse Spark model page, read September 26, 2026. All versions on a tier cost the same, so upgrading from 1.2 to 1.3 doesn't change the rate. Reasoning tokens are billed as output. Web search grounding costs an extra $2.50 per 1,000 searches.

The Contributor discount comes with a condition. Meta says it can use your prompts and completions "to train future Meta models." Standard-tier prompts and completions aren't used for training. So don't send data to a Contributor model unless you're allowed to share it with Meta. Contributor also has lower rate limits and can't use max reasoning on the API.

Rate limits, API keys, and billing belong to Meta Model API, not to the model. We don't cover them here.

Where you can use Muse Spark

These surfaces don't expose the same controls. The API limits max reasoning to Standard 1.3, and Muse Code's docs say otherwise. We haven't checked which reasoning levels OpenRouter passes through, and in the consumer apps you can't pick a version at all.

OpenCode Go, a third-party subscription for coding agents, also includes Muse Spark 1.3 Contributor and 1.2 Contributor, with Meta's Contributor training terms attached.

Muse Code: Meta's terminal coding agent built on Muse Spark, with its plans and platform support

entity↗

Meta Muse: the consumer agent that runs on Muse Spark

entity↗

OpenCode Go: Muse Spark Contributor models inside a flat-rate coding-model subscription

product↗

Is Muse Spark open source or open-weight?

No. As of September 26, 2026, Meta only offers Muse Spark as a hosted model. There are no downloadable weights, and you can't self-host it. Meta's Models docs list Muse Spark among the models "hosted on Meta Model API" and name Muse Glimmer as the one that runs on your own hardware.

Muse Glimmer is where the confusion comes from. Meta released Glimmer on August 10, 2026, under the Apache 2.0 license. It's a 30B dense multimodal model distilled from Muse Spark's outputs. It takes text and images, has a 128K default context, and is built to run on a single 24 to 32 GB GPU. Some pages attach "Apache 2.0" or "open-weight" to Muse Spark, but that license covers Glimmer only. If you need a Meta model you can self-host today, it's Glimmer, not Spark.

Meta's benchmark claims

These numbers are vendor claims. Meta ran them, and we haven't reproduced them.

Taken at face value, Meta's own table has 1.3 ahead of the other models on long-context retrieval, DeepSWE, and SWEAtlas, tied with GPT 5.6 Sol on Terminal-Bench 2.1, and behind Opus 5 on GDPVal-AA, OSWorld 2.0, and AutomationBench. That's what Meta reports, not proof that Muse Spark is better.

Meta's pages disagree on some scores

Meta's August 5 launch post for 1.2 has its own bar charts, and each model there is labeled with the harness it ran in. Several of those numbers don't match the 1.3 page:

The launch charts also compare against GPT 5.6 Terra (in Codex), while the 1.3 page uses GPT 5.6 Sol. Those are different models, so don't read one as the other.

What outside evaluations say

Outside results are thinner than Meta's charts, and they don't all agree.

Taken together, the outside evidence supports a narrow claim: Muse Spark 1.3 is competitive on aggregate benchmarks, and on those benchmarks it uses a lot of tokens. It doesn't settle whether 1.3 is better than GPT, Claude, or Gemini models at real work. We don't make that call here.

How to evaluate a model on your own task instead of trusting a leaderboard

concept↗

Muse Spark vs other things called "Muse"

What we haven't verified

• Which Muse Spark version meta.ai and the Meta AI app run today.
• When Muse Spark weights will ship, which version, and under what license. Meta's roadmap line and the reported 1.2 pledge are all there is.
• Whether max reasoning works on Contributor models in Muse Code, as Muse Code's docs say, when the API docs say it doesn't.
• Which reasoning levels OpenRouter supports, and whether Meta's audio caveat for 1.3 applies there too. OpenRouter lists audio input for 1.3.
• Why Meta's DeepSWE numbers differ between its pages.
• How Muse Spark does on real builder tasks. We haven't run it.

How we separate vendor claims, outside reports, and our own measurements

concept↗

Change log

Sources & Documentation

1. Meta: Muse Spark 1.3 model page — 1.3 benchmark table, 1.3 Standard and Contributor prices, 1M context, availability surfaces.
2. Meta Model API docs: Models — Five model IDs, modalities, 1,048,576-token context, 1.3 audio caveat, max on Standard only, Muse Glimmer is self-hosted.
3. Meta Model API docs: Pricing and rate limits — Standard and Contributor rates, training-data tradeoff, no long-context premium, search grounding price.
4. Meta Model API docs: Reasoning — Reasoning effort levels. none unsupported. max limited to Standard-tier muse-spark-1.3.
5. Meta docs: Muse Code configuration — Default model muse-spark-1.2. Muse Code offers max on both Standard and Contributor.
6. Meta: Meta Model API product page — Public preview, all five Muse Spark model IDs and prices, OpenRouter availability.
7. Meta AI blog: Introducing Muse Spark (April 8, 2026) — Original launch in meta.ai and the Meta AI app, private API preview, Contemplating mode.
8. Meta AI blog: Introducing Muse Spark 1.1 (July 9, 2026) — 1.1 release, Meta Model API public preview, Thinking mode in the Meta AI app, 1M context.
9. Meta AI Research: Introducing Muse Code and Muse Spark 1.2 (August 5, 2026) — 1.2 release, co-training with Muse Code, Terminal-Bench 2.1 and DeepSWE 1.1 charts with harness labels.
10. Meta AI Research: Introducing Muse Spark 1.3 (September 2, 2026) — 1.3 release, max reasoning, efficiency claim, open weights on the roadmap.
11. Meta AI Research: Introducing Muse Glimmer (August 10, 2026) — Glimmer open weights under Apache 2.0, distilled from Muse Spark.
12. Meta Model API docs: Muse Glimmer — 30B dense multimodal model, 128K default context, Apache 2.0 license.
13. OpenRouter: Muse Spark 1.3 — Listing and prices for Meta's Muse models on OpenRouter, checked via OpenRouter's models API.
14. Constellation Research: Meta releases open weight Muse Glimmer model with open Muse Spark 1.2 on tap — Press report of the Muse Spark 1.2 open-weight pledge.
15. Wikipedia: Muse Spark — Cited for its July 9 release date and its CNBC-sourced account of the 1.2 open-weight pledge.
16. Artificial Analysis: Muse Spark 1.3 (max) — Independent Intelligence Index score, token use, speed, and cost per task.
17. MindStudio: Meta Muse Spark 1.3 benchmark confusion (September 5, 2026) — Single hands-on coding comparison and leaderboard observation.