Muse is the personal agent for regular people. Muse Spark is the closed, API-only AI model that powers it and Meta AI. Muse Glimmer is a smaller open-weight model (30B, Apache 2.0) you can run yourself. Muse Image makes pictures inside Instagram and WhatsApp. Muse Code is a terminal coding agent for developers. Muse Voice Transcribe turns speech into text. Muse Charm is a wearable device for the agent. Everything except the agent is aimed at developers or is a feature inside another Meta app.
The family at a glance
| Name | Type | For | Open weights? | How you get it |
|---|---|---|---|---|
| Muse | Personal AI agent | Everyone | n/a | Muse app, muse.ai, WhatsApp, Mac |
| Muse Spark 1.3 | Frontier reasoning model | Muse itself, Meta AI, developers | No, API only | Meta Model API, OpenRouter, Vercel AI Gateway, Bedrock |
| Muse Spark Contributor | Discounted version of Spark | Developers who allow training on their data | No | Meta Model API, OpenRouter |
| Muse Glimmer 30B | Open-weight model | Developers, hobbyists, on-device | Yes, Apache 2.0 | Hugging Face, Ollama, Unsloth, OpenRouter |
| Muse Image 1.0 | Image generation model | Instagram and WhatsApp users; developers | No | Meta AI in apps; API, priced per image |
| Muse Code | Terminal coding agent | Programmers | Tool is downloadable; model is Spark | Install from Meta's developer site |
| Muse Voice Transcribe 1.0 | Speech-to-text model | Developers | No | API, priced per audio second |
| Muse Realtime Avatar | Model for live video avatars | Muse users (coming) | No | Inside Muse |
| Muse Charm | Wearable device | Muse users | n/a | Pre-announced by Meta |
Muse, the agent
The consumer product this whole directory is about. Launched September 8, 2026 in the US, later Canada. It runs on Muse Spark, has its own cloud computer with a browser, connects to your apps, and does tasks for you. If you are not a developer, this is the only entry on this page you will use directly. Start with What is Meta Muse?
Muse Spark 1.1, 1.2, 1.3
Muse Spark is the large language model built by Meta Superintelligence Labs (MSL), the AI division Meta reorganised in mid-2025 under chief AI officer Alexandr Wang. It is the first model in the Muse family and the successor, in practice, to Llama as Meta's flagship, though Llama continues as an open-weight line.
Versions.
- Muse Spark (announced April 8, 2026): MSL's first model, introduced as "purpose-built to prioritize people."
- Muse Spark 1.1 (July 9, 2026): the first public release. Multimodal reasoning, coding, up to a million tokens of context. Also the moment Meta AI switched to running on it. Launched with the public preview of the Meta Model API for US developers.
- Muse Spark 1.2 (August 5, 2026): a reasoning model "designed for complex agentic tasks." Released alongside Muse Code and made available on the Meta Model API and OpenRouter. Zuckerberg has said Meta intends to release 1.2 as open weights; as of this writing it has not.
- Muse Spark 1.3 (September 2, 2026): the current version and the one behind the Muse agent. Multimodal (text, images, video, PDFs, audio in; text out), a 1,048,576-token context window, trained for long-running agentic and multi-agent workflows with reliable tool calling and "competitive coding performance." Meta notes that audio understanding in 1.3 is not fully supported and recommends 1.2 or Muse Voice Transcribe for audio.
Pricing. Through gateways, Muse Spark 1.3 is listed at about $1.25 per million input tokens and $4.25 per million output tokens, with cached-input discounts. The "contributor" variants (muse-spark-1.3-contributor, 1.2-contributor) are heavily discounted in exchange for permission to use your prompts and completions to train future models. Standard versions are not used for training.
Is it open source? No. Muse Spark is closed and served over an API only. This is the most common confusion about it, because Meta's Llama models were open and because Glimmer, below, is. If you need weights you can download, that is Glimmer.
How good is it? Meta publishes headline benchmark scores as an image on its release post. Independent trackers say 1.3 closed much of the gap on overall intelligence indexes, edging some rivals on aggregate, while ChatGPT and Claude remain ahead on coding. Treat every benchmark table, including ones on this site, as directional.
Muse Glimmer
Released August 10, 2026, Muse Glimmer is the open one. It is a dense, 30-billion-parameter multimodal model distilled from Muse Spark and "optimized for autonomous agents on consumer hardware." Apache 2.0 licence, 128K default context, image understanding, tool use with failure recovery, and support for more than 100 languages.
The point of Glimmer is that it runs on a single good computer. Meta's own framing: "an open agentic model that runs on your device." It is on Hugging Face as meta-models/Muse-Glimmer-30B, available through Ollama, in Unsloth fine-tuning builds, and on OpenRouter for hosted use. People compare it against Qwen's 27B and 35B models, Google's Gemma 4 and Nvidia's Nemotron, which is the right weight class.
Glimmer does not power the Muse agent. It is for developers who want a permissively licensed model they control, including a self-hosting escape hatch that Spark does not offer.
Muse Image
Meta's image generation model, launched July 7, 2026, and marketed as "the creative partner that knows your world." It appears inside Instagram and WhatsApp as a Meta AI feature, with Facebook and Messenger to follow, and it is available to developers as muse-image-1.0, priced per image rather than per token, with its own rate limit.
Its launch is the reason "muse image privacy" and "instagram muse opt out" are big searches. A feature let anyone tag a public Instagram account in a prompt and generate images from that person's photos, on by default, without notifying them. Meta pulled the tagging feature within days after objections from CAA and SAG-AFTRA, but the underlying "sharing and reuse" setting remains on for public accounts. The exact steps to turn it off are in our privacy guide. Muse Image is a separate product from the Muse agent.
Muse Code
Released August 5, 2026 alongside Spark 1.2, Muse Code is a terminal-based coding agent: "multi-agent coding in your terminal, built for Muse Spark." It can write and debug code, work through longer software tasks, and run multiple sub-agents at once. It competes with tools like Claude Code, OpenAI's Codex CLI and OpenCode.
Pricing follows the Meta Model API's token pricing, and Meta's developer site points to a GitHub-agent cookbook that combines Muse Code, the Model API and OpenCode to build an autonomous GitHub bot. Whether it has a free tier has changed over time; check Meta's developer site rather than trusting a cached answer. Muse Code is for programmers; it has nothing to do with the consumer Muse app beyond the shared model.
Muse Voice Transcribe
A synchronous speech-to-text model (muse-voice-transcribe-1.0) for push-to-talk, endpointing and speaker-aware transcription, with keyword biasing for domain terms and language hints. Priced by audio second on Meta's API and available through OpenRouter. Developers use it; consumers meet it indirectly when they talk to Muse or Meta AI.
Muse Realtime Avatar
Announced at Meta Connect on September 23, 2026. A model that gives your Muse a face, body and voice so you can video chat with it live, hand it tasks and ask questions. Zuckerberg said the physical avatar is named Jolly. Coming soon inside the Muse app.
Muse Charm
Also from Connect: a small wearable device Meta describes as "the first device made for Muse," letting you reach your agent by voice, touch and camera. TechCrunch called it Tamagotchi-like. Meta has a sign-up page for launch notifications; no price or date yet. Separately, Muse is coming to Meta's AI glasses with a wake word "in the coming months."
Which are open source?
- Open weights: Muse Glimmer 30B (Apache 2.0). That is the list.
- Closed, API only: Muse Spark (all versions, including contributor tiers), Muse Image, Muse Voice Transcribe, Muse Realtime Avatar.
- Promised: Zuckerberg has said Spark 1.2 will be released as open weights. Not yet delivered.
- Tools: Muse Code is a downloadable tool, but it runs on the closed Spark model.
If a page tells you "Muse Spark is open source," it is either wrong or talking about Glimmer.
How developers get access
- Meta Model API (dev.meta.ai): direct, self-serve, pay-as-you-go access to Spark, Image, Voice Transcribe and SAM 3.1. Rate limits by tier; teams share quotas across keys.
- OpenRouter: all the Spark versions, contributor tiers, Glimmer and Voice Transcribe, so you can swap Muse models into existing tools.
- Vercel AI Gateway, Amazon Bedrock and others: Spark 1.3 is listed on several gateways.
- Hugging Face, Ollama, Unsloth: Glimmer weights for local or fine-tuned use.
- Muse connectors: if you want your app to work inside the Muse agent, Meta opened the connector platform in September and received more than 1,500 applications in its first week. The agent uses the same open standard (MCP) that other AI tools use. Our page for app makers covers submission and the live status badge.
Timeline
- April 8, 2026: Muse Spark announced, MSL's first model.
- July 7: Muse Image launches in Instagram and WhatsApp; tagging feature pulled within days.
- July 9: Muse Spark 1.1 released; Meta AI switches to it; Model API public preview.
- August 5: Muse Spark 1.2 and Muse Code.
- August 10: Muse Glimmer 30B, open weights.
- August 19: Meta AI Mac app (beta).
- September 2: Muse Spark 1.3.
- September 8: Muse, the personal agent, launches in the US.
- September 16: Invite codes, 1 billion tokens each.
- September 18: Muse for Mac.
- September 23: Meta Connect: Realtime Avatar, glasses integration, Mac computer control, Muse email address, retail partners, Muse Charm.
We update this page as the family grows. For the agent itself, start at Start here.