Meta Muse Spark explained: the closed model behind Meta AI, the Muse agent app, and what's open

Update, October 22, 2026: On October 20, 2026, Meta paused sharing in Muse after researchers found more than 6,800 share links in DNS records; see what was exposed. Anthropic's comparable agent, Claude Duet, is covered in our explainer.

Update, October 16, 2026: On October 15, Meta released Muse Spark Open, an 816B-parameter open-weight model (42B active) under its new Muse Model License, so Muse Glimmer is no longer the only open-weight Muse model. Meta has since also launched the closed Muse Wonder models. See Muse Spark Open: weights, license and hardware and Muse Wonder.

Muse Spark is the family of frontier models that Meta Superintelligence Labs (MSL) has shipped since April 8, 2026. It runs Meta AI, Meta's coding agent Muse Code, and Muse, the personal agent app Meta launched in the US on September 8. On October 7, Muse got its own iPad app. Muse Spark is also a break with Meta's past. Unlike Llama, its weights aren't published, and six months on, the only open-weight model in the Muse line is a smaller distilled one.

This explainer is based on Meta's announcements, its developer site and Hugging Face page, and reporting from TechCrunch, The Register, AP and the Wall Street Journal, checked October 12, 2026.

The short version

• What: Meta's multimodal reasoning model family. The current version is Muse Spark 1.3, released September 2, 2026.
• Open or closed: closed weights. You use it through Meta's apps or its paid API.
• API price (Muse Spark 1.3): $1.25 per million input tokens and $4.25 per million output, with a 1-million-token context window. A much cheaper "contributor" tier lets Meta use your data to improve its products.
• The Muse app: a personal agent that can browse, fill in forms, send email and buy things. It's US-only, free, with Power ($20/month) and Maximum ($100/month) plans, according to TechCrunch.
• What's open: Muse Glimmer, a 30-billion-parameter model under Apache 2.0, released August 10. Open Muse Spark weights are promised but haven't been released.

What is Muse Spark?

Meta introduced Muse Spark on April 8, 2026, as "the first in a new series of large language models built by Meta Superintelligence Labs." Its technical post calls it "a natively multimodal reasoning model with support for tool-use, visual chain of thought, and multi-agent orchestration." Meta didn't publish a parameter count or architecture details.

Meta made three claims at launch. All of them are vendor-reported and haven't been independently checked:

• Efficiency. Meta says a rebuilt pre-training stack reaches the same capabilities as its previous model, Llama 4 Maverick, with "over an order of magnitude less compute."
• Reasoning. In a "Contemplating" mode that runs several agents in parallel, Meta reports 58% on Humanity's Last Exam and 38% on FrontierScience Research, and says that mode competes with Gemini Deep Think and GPT Pro. Meta didn't publish the full settings behind those scores in the text of its post.
• Weak spots. Meta itself said gaps remained in "long-horizon agentic systems and coding."

In the Meta AI app, Muse Spark launched with Instant and Thinking modes, and Contemplating rolled out gradually on meta.ai. If the modes are unfamiliar, our explainer on reasoning models covers what "thinking" costs and buys.

How has it changed since April?

Meta has shipped point releases roughly every month:

Meta's September 2 post says Muse Spark 1.3 "with max reasoning is now available on Muse Code and Meta Model API." Meta also says 1.3 is better at asking clarifying questions, confirming before "consequential actions," and resisting prompt injection. Its comparison charts against GPT 5.6 Sol and Claude Opus 5 are vendor-run, and the scores appear only as images.

Is Muse Spark open source?

No. This is the biggest change from the Llama era. In 2024, Mark Zuckerberg wrote that "open source AI represents the world's best shot at harnessing this technology," and Meta published Llama's weights under its own license. Muse Spark's weights have never been published. The Register's launch report summed it up: "no Llama." Meta's April newsroom post said only that "we hope to open-source future versions of the model."

Meta has partly moved back toward openness since then:

• Muse Glimmer (August 10). Meta's research post describes "a 30-billion-parameter model" released "under a permissive Apache 2.0 license," small enough to run on "a Mac or PC with a single consumer GPU" once quantized. It was trained partly "on Muse Spark's outputs using logit distillation." It is a separate, smaller model, not Muse Spark's weights.
• Muse Spark 1.2 (promised). The Wall Street Journal reported on August 10 that Meta planned to release an open-weight version of Muse Spark 1.2 "in the coming weeks." As of October 12, Meta's Hugging Face organization lists only Muse Glimmer models, and Meta's September 2 post lists "the Muse Spark open weights release" on its roadmap without a date.

So today the line splits in two: a closed flagship, and a small open model distilled from it. Llama hasn't been updated since Llama 4, so the license terms in our comparison of open-weight licenses still apply to Llama, while Muse Glimmer's Apache 2.0 license is one of the permissive ones. For whether running a model yourself saves money, see frontier APIs vs open models: the real cost.

How much does the Muse Spark API cost?

Meta's developer site lists two tiers for Muse Spark 1.3:

Last verified: October 12, 2026 (dev.meta.ai model page)

The contributor tier is about 12 times cheaper on input and 21 times cheaper on output. In exchange, Meta gets to keep and use what you send. That's a real trade-off for anyone handling customer or company data. Meta's developer site also lists Muse Spark 1.1 and 1.2, Muse Glimmer, an image model ($0.01 per image) and a speech-to-text model.

What is the Muse app?

Muse is Meta's consumer agent: software that acts for you, not just answers. Our explainer on what "agentic" really means covers the distinction. Meta announced it on September 8, 2026.

What it does, according to Meta: sends email, books travel, fills out forms, opens a browser and "negotiate[s]" on your behalf. It turns long-term goals into plans, keeps working after you close the app, and comes back when something changes or needs your approval. AP reports it is for adults 18 and older.

Where it runs: in the US on iOS, Android, the web at muse.ai, and inside WhatsApp, with Meta's AI glasses "coming soon." On October 7, TechCrunch reported a dedicated iPad app and new connectors for Asana, Canva, Dropbox, Figma, QuickBooks, GitHub, Klaviyo, Zoom, Notion, Granola and Meta ad accounts. It also reported shopping partners including Best Buy, Gap, Sephora, Walmart and Wayfair. TechCrunch cited Sensor Tower estimates of more than 6.6 million installs in the first month.

What it costs: Meta's post says Muse is free "for most everyday needs," with subscriptions for heavier use. TechCrunch reported the two paid plans, Power at $20 a month and Maximum at $100 a month, and said a payment card is required to start.

How Meta says it protects you:

• Muse runs on a dedicated cloud machine Meta calls Muse Secure VM.
• A separate Sentinel agent must approve anything Muse sends to the internet.
• Muse can use saved credentials without seeing passwords or card numbers. Checkout goes through Link by Stripe, which issues one-time card numbers.
• Muse asks before sensitive actions such as sending email or buying, and keeps an audit trail.
• Conversations aren't shared with Meta's ad systems, and you can opt out of training.

Two caveats. TechRepublic, citing WIRED, reports that Meta acknowledges the Secure VM isn't technically inaccessible to Meta, only off-limits by policy. A stricter Confidential VM, encrypted with a key only you hold, is promised "later this year." And an agent that reads your email and browses the web can be manipulated by instructions planted in what it reads. Approval prompts reduce that risk but don't remove it.

How does Muse compare with other agents?

Muse arrived alongside several workplace and consumer agents. Google announced its Gemini agent for work on October 8, and, according to CBS News, OpenAI launched its "Dots" helper for ChatGPT Pro and Business Premium users in late September. Muse differs in two ways. It's aimed at consumers rather than IT departments, and it rides on WhatsApp's reach. On October 6, according to TechCrunch, Meta also announced work with partners on an open standard that would let personal agents identify themselves to websites, which sometimes block them as bots.

What's confirmed, reported and uncertain

• Confirmed (Meta): the April 8 launch and closed weights; the Muse Spark 1.3 release on September 2; API prices and the 1M context; the Muse launch on September 8, with its platforms and security design; Muse Glimmer's size and Apache 2.0 license; an open-weights Muse Spark release on the roadmap.
• Vendor-reported: the compute-efficiency claim, the benchmark scores, and the 1.3 efficiency gains.
• Reported: Muse's $20 and $100 plans (TechCrunch); the iPad app, connectors and install estimates (TechCrunch, Sensor Tower); the plan to open Muse Spark 1.2 (WSJ); Meta's access to the Secure VM (WIRED via TechRepublic); the 1.1 release date (CNBC, TechCrunch).
• Uncertain: when, or whether, Muse Spark weights will be published, and under which license; Muse availability outside the US; when the Confidential VM ships.