Muse Spark 1.3 on the Meta Model API

Last reviewed 19 September 2026. Every claim below is backed by a source in Sources. Where we have no source, the guide says so.

First, the name

There is no "Muse Spark API". Muse Spark 1.3 is a model. The API it is served through is the Meta Model API, at dev.meta.ai, which serves Meta's models generally.

We had this wrong ourselves at first and corrected it, which is why the distinction gets a section rather than a footnote. It matters when you are searching for docs: you are looking for a model id inside a general-purpose API, not for a product-specific endpoint.

What shipped, and when

Meta Research announced Muse Spark 1.3 on 2 September 2026, describing it as "now available on Muse Code and Meta Model API". That is six days before the Muse assistant itself debuted, on 8 September.

The two model ids

Meta's developer docs list two:

• muse-spark-1.3
• muse-spark-1.3-contributor

We have not found a published explanation of what separates the contributor variant from the base model, so we are not going to guess at one. If you know, and can point at a page that says so, send it over and we will add it with the source.

The context window

1,048,576 tokens — exactly 2²⁰, which is the usual way a "1M context" window is actually specified.

That is the figure in the docs. We have not tested retrieval quality across that window and have not seen a published evaluation of it, so treat it as the stated capacity rather than a promise about behaviour at the far end of it.

Where you can use it

Two surfaces, both named in the announcement:

• The Meta Model API — the general developer API at dev.meta.ai, where you select the model by id.
• Muse Code — Meta's coding tool, which is covered in its own guide.

One practical note for anyone trying to follow this: Meta's developer documentation site is publicly crawlable — its robots.txt is permissive and its sitemap is substantial. That is the opposite of the Muse consumer site, where automated collection is prohibited. If you are looking for primary sources on the model, the developer docs are the place they exist.

How this relates to Muse connectors

Mostly, it doesn't — and conflating the two is the most common mistake we see.

A Muse connector is a consumer-facing integration: you link a product to the Muse assistant, and Muse acts in it on your behalf. Submissions go through muse.ai/platform, and there is a guide to that form.

Muse Spark 1.3 is a model you call yourself, with your own key, from your own code. Building on Spark does not put anything in the Muse directory, and shipping a connector does not require you to use Spark for anything.

We have not seen Meta publish which model powers the Muse assistant, so we do not claim it is this one.

What is not published

Things people ask that we cannot answer from a source:

• Pricing, rate limits and quotas.
• Benchmark results, and how the model compares to anything else.
• Which modalities it takes and returns.
• Maximum output length, as distinct from the context window.
• What the -contributor variant is for.
• Whether the Muse assistant runs on it.