Models/meta/Muse Spark

Proprietary (Meta). API access through the Meta Model API and OpenRouter; weights not released

Muse Spark

Name: Muse Spark
Author: meta

by meta

Meta's proprietary frontier model: text, image, video, audio and PDF in, text out, one million tokens of context

Not disclosedparams

Source Run it on your own data, free

Identifiers

Model ID

meta/Muse Spark

Feature URI

Overview

Muse Spark is the first model in the Muse family from Meta Superintelligence Labs. Meta announced it on April 8, 2026 as the model behind Meta AI, opened it to developers through the Meta Model API as Muse Spark 1.1 on July 9, 2026, shipped 1.2 alongside the Muse Code terminal agent on August 5, 2026, and released Muse Spark 1.3 on September 2, 2026. It takes text, images, video, audio and PDF documents as input and returns text, with a context window of one million tokens, Instant and Thinking modes, tool calling with parallel calls, structured output, and the ability to launch sub-agents. Meta positions it for agentic workflows and long-horizon coding. The weights are not released; Meta has said it hopes to open-source future versions, and the open-weight sibling is Muse Glimmer, a 30-billion-parameter model under Apache 2.0.

Architecture

Meta has not published the parameter count or the architecture. What is public: natively multimodal perception over video, images and documents, a 1M-token context window, two inference modes (Instant, Thinking), and an OpenAI-compatible API. Muse Spark 1.3 is served under two identifiers on the Meta Model API: muse-spark-1.3 and muse-spark-1.3-contributor, the latter priced lower in exchange for the traffic being used to improve Meta's products.

Mixpeek SDK Integration

import { Mixpeek } from "mixpeek";

const mixpeek = new Mixpeek({ apiKey: process.env.MIXPEEK_API_KEY });

// Muse Spark is a reasoning model, not an embedding model, and Mixpeek does not
// run it. Index with a hosted extractor; Muse Spark then reasons over what the
// retriever returns (see the retrieve sample).
await mixpeek.collections.create({
  collection_name: "media_library",
  feature_extractors: [
    { feature_extractor_name: "multimodal_extractor", version: "v2" },
  ],
});

Capabilities

Multimodal input: text, images, video, audio and PDF documents in one prompt
One million tokens of context
Instant and Thinking modes, tool calling with parallel calls, structured output
Sub-agent orchestration for long-horizon agentic and coding tasks
OpenAI-compatible API on the Meta Model API and through OpenRouter

Use Cases on Mixpeek

Reasoning over retrieved video segments, transcripts and documents in one answer

Agentic workflows that read a Mixpeek retriever's results and decide the next call

Drafting structured descriptions or labels from multimodal evidence a retriever surfaces

Coding agents that need long context over documentation and source

Common Pipeline Companions

google/gemini-embedding-2

Multimodal embeddings for the retrieval step Muse Spark reasons over

openai/whisper-large-v3

Transcripts that give Muse Spark the spoken words next to the frames

Frequently Asked Questions

Does Mixpeek run Muse Spark?

No. Muse Spark is API-only from Meta, and Mixpeek's hosted extractors and the LLM stages in retrievers (llm_enrich, summarize) do not list Meta as a provider. Use Mixpeek to index and retrieve, then call Muse Spark with the retrieved results, either from your application or from a retriever's api_call stage.

Are the weights open?

No. Muse Spark is proprietary and served through the Meta Model API and OpenRouter. Meta has said it hopes to open-source future versions. The open-weight model in the same family is Muse Glimmer, 30 billion parameters under Apache 2.0, released August 10, 2026.

What does it take as input?

Text, images, video, audio and PDF documents, with a context window of one million tokens. Output is text, including structured output and tool calls.

What is the difference between muse-spark-1.3 and muse-spark-1.3-contributor?

Same model. The contributor identifier is priced far lower on the Meta Model API in exchange for the traffic being used to improve Meta's products; the plain identifier is not used that way. Pick by your data-handling requirements, not by capability.