Muse is Meta's personal AI agent, and it moves fast: Muse Wonder, Muse Code, the open Muse Spark models and new hardware like Muse Charm all landed within weeks. Here are the Muse AI features, use cases and tips worth knowing, plus the news and privacy questions around them. Every item links the original post or report.
17 items · 11 new this week · updated Oct 15, 2026
Nothing matches that filter yet. Try another topic.
Muse's own account suggested a timely trick: ask your Muse "based on my IG likes from this year, what should I be for halloween?" It shows off what makes Muse different, since it can draw on your connected Meta accounts for personal suggestions.
Why it matters: It shows the personal side of Muse: suggestions based on your own Meta activity rather than generic lists.
Good for:
Muse users who have connected Instagram.
Plan & availability:
Muse app and web; needs Instagram connected to your Muse.
Cursor added Muse Wonder from Meta Superintelligence Labs on October 15. Cursor says it scores 46.1% on CursorBench at default effort, the highest of any Muse model, and it is priced at $4/M input and $20/M output tokens. You can try Meta's model without leaving your editor.
Why it matters: You can test Meta's newest model on your real codebase from the editor you already use.
Good for:
Cursor users comparing models for agent work.
Plan & availability:
Cursor model picker; billed at $4/M input and $20/M output tokens per Cursor.
Meta released Muse Spark Open on October 15 and calls it the most capable US open-weight model to date: 816B total and 42B active parameters, with a 1M-token context. It is built for coding and agent work, runs on two Apple M5 Ultra Mac Studios with MLX, and is available on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
Why it matters: A large open-weight model with 1M context gives teams a frontier-class option they can host and control themselves.
Good for:
ML engineers, startups that need self-hosting, and open-model fans.
Plan & availability:
Weights on Hugging Face; hosted on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Check the license.
Hugging Face's launch post shows how to run Muse Spark Open on Apple silicon through llama.cpp or MLX. It is a very large model, so even the 3-bit MLX quant needs more than 256 GB of RAM. Hugging Face reports about 21 tokens a second on an M3 Ultra with 512 GB. Reasoning can't be switched off, so set it to low for easy tasks to save tokens.
Why it matters: It tells you up front what hardware you really need, before you spend a day downloading weights.
Good for:
Mac users with high-memory Apple silicon who want to run models locally.
Plan & availability:
Free open weights; needs well over 256 GB of RAM even at 3-bit.
Meta built a human-led research setup powered by Muse Spark 1.3 and Muse Wonder, with access to 5 million de-identified health records and public data. Medical experts used it for hypothesis generation, code and manuscript drafts while keeping control of every scientific decision. Meta shared 10 resulting manuscripts across cardiology, oncology, infectious disease, emergency medicine and more.
Why it matters: It's a concrete example of an AI-assisted research workflow where experts stay in charge of the judgement calls.
Good for:
Clinical researchers, health data scientists and research leads.
Plan & availability:
Research project from Meta; the setup itself isn't a public product.
Meta's AI account says Muse Wonder and Muse Wonder Full reach state-of-the-art results on personal and professional agentic tasks, and links a researcher blog with the details. Meta's chief AI officer, Alexandr Wang, separately said Wonder Full topped a new coding benchmark released the same day.
Why it matters: Vendor benchmarks are a starting point. They tell you which jobs to test the model on yourself.
Good for:
Developers and teams comparing frontier models for agents.
Applies to:
Muse Wonder in Muse paid plans, Muse Code and the Meta Model API.
Alexandr Wang, Meta's chief AI officer, shared an example of an advanced physics visualisation built with Muse Wonder Full. It's a good hint at a strength to test: turning a physics idea into an interactive, animated visual.
Why it matters: Interactive visuals make hard physics ideas click far faster than equations on a page.
Good for:
Teachers, students and science communicators.
Plan & availability:
Muse Wonder Full is on Muse Code's Power and Maximum plans.
Meta launched Muse Wonder, its new frontier model family, on October 12. In the Muse app, Wonder powers the assistant for subscribers on the $20 Power or $100 Maximum plan, and the free tier stays on Muse Spark 1.3. Meta is rolling it out in waves, so it may not show up right away.
Why it matters: Paying Muse users get Meta's strongest model for multi-step errands; free users stay on Spark 1.3.
Good for:
Muse users deciding whether to upgrade.
Plan & availability:
Muse Power ($20) and Maximum ($100) plans; rolling out in waves.
For developers, Muse Wonder is available in Muse Code and the Meta Model API at $4 input and $20 output per million tokens. Muse Wonder Full is only available on Muse Code's Power and Maximum subscription plans. Meta pitches both for frontier coding, long-running agents and computer use.
Why it matters: Developers can use Muse Wonder in Meta's own coding agent or through the API at a published price.
Good for:
Developers evaluating Meta's models for coding agents.
Plan & availability:
Muse Code and Meta Model API ($4/$20 per M tokens). Wonder Full only on Muse Code Power and Maximum.
On launch day, a developer gave Muse Code (Muse Wonder, extra-high effort) and Claude Code (Opus 5.5, extra-high effort) the same one-shot job: a playable first-person 3D game of his kitchen in a single three.js file. He reports both ran first try; Opus took 11m40s and about $2.10, Wonder took 3m42s and about $1.
Why it matters: Side-by-side one-prompt tests are a cheap way to compare speed and cost on the kind of work you actually do.
Good for:
Developers deciding whether to try Muse Code or Muse Wonder.
Plan & availability:
Muse Code with Muse Wonder through the API; Wonder Full needs a larger Muse plan.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
A community benchmark that has AI models play Pokémon Red started a live run with Muse Wonder on launch day. The account says Muse Spark 1.3 previously used all 1,000 turns and only reached Viridian Forest, and that the record is held by GPT-6.1 Sol at 244 turns, with the same prompt and setup for every run.
Why it matters: Long game runs are a fun, public way to see how well a model plans over many steps.
Good for:
Anyone curious how agent models compare on long tasks.
Plan & availability:
Community benchmark; results aren't from Meta.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
Meta brought Muse to iPad on October 7, a month after the phone launch. TechCrunch reports that new connectors include Asana, Canva, Dropbox, Figma, QuickBooks, GitHub, Klaviyo, Zoom, Notion, Granola and Meta ad accounts. With those, Muse can work across your business tools, not just email and calendar.
Why it matters: New connectors let Muse work across real business tools, not just email and calendar.
WIRED reported that Muse's operating instructions, which researchers got the agent to export, describe an hourly process that builds a page for each person in a user's life, such as family, friends and colleagues, with sections like shared history and important dates. Meta told WIRED that an agent needs context about you and the people you deal with to help you, that it uses public information and what you choose to share, and that you can edit or wipe memories and disconnect services.
Why it matters: If you connect messages, photos or accounts to Muse, it's worth knowing how it may organise information about other people.
Good for:
Muse users, especially those who have connected messaging or social accounts.
Applies to:
Applies to the Muse app; settings and behaviour may change.
A creator shared a prompt that turns Muse into a Meta ads specialist, and argued it can replace some of the work small businesses pay agencies for. Muse can connect to Meta ad accounts, so it can look at your real campaign data rather than giving generic advice.
Why it matters: Small advertisers can get a first-pass audit of their own campaigns without hiring someone.
Good for:
Small business owners and solo marketers who run Meta ads.
Plan & availability:
Muse app with your Meta ad account connected.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
A marketer described a weekly routine: connect Muse to the tools that each see part of the business (Shopify for sales, Stripe for payments, QuickBooks for spending, Meta Ads for demand, Klaviyo for email, Slack and Notion for plans), then give it business questions instead of tool-by-tool ones. His starter prompt asks Muse to explain what moved revenue this week, show the evidence, and turn it into the next campaign brief.
Why it matters: Answers that need data from five tools usually take an afternoon of exports. One connected agent can trace them in one go.
Good for:
Ecommerce founders and small teams with their tools spread across several apps.
Plan & availability:
Muse with the relevant connectors; which connectors are available depends on your plan and region.
Note: reported by users on X; we haven't independently confirmed it works in the current version.
Muse Gadgets is Meta's open-source project for building your own hardware that talks to Muse. You can program an off-the-shelf ESP32 board or turn a Raspberry Pi or other Linux box into a gadget, then connect Muse to screens, buttons, sensors and actuators. The SDKs and firmware are Apache 2.0. TechCrunch reports project ideas such as a colour e-ink display and an HDMI stick for a TV.
Why it matters: Muse can now live outside your phone on hardware you design, which is a big deal for makers and home-automation fans.
Good for:
Makers, Raspberry Pi and ESP32 hobbyists, and Home Assistant users.
Plan & availability:
Free and open source; you need a Muse account and an SDK token to pair a gadget.
At Meta Connect, Mark Zuckerberg briefly showed the Muse Charm, a palm-sized device just for the Muse agent. The Verge describes a big screen, a fingerprint sensor you press to talk, a small camera and microphones. Zuckerberg said you won't need to unlock a phone to use it and that Meta plans to ship it in time for the December holidays. Price wasn't announced.
Why it matters: It's Meta's bet that a personal agent deserves its own device, alongside its glasses.
Good for:
Early adopters and anyone deciding whether to wait for Muse hardware.
The latest we have covered: Ask your Muse for a costume idea based on your Instagram likes (Oct 15); Use Muse Wonder inside Cursor (Oct 15); Run Meta's open-weight Muse Spark Open yourself (Oct 15). Every item on this page links its original source.
Muse is Meta's personal AI agent. It launched in September 2026 and works in the Muse app, on WhatsApp, on the web and on Mac and iPad. It can connect to your apps and accounts and carry out tasks for you, asking before actions like sending or buying.
Muse Wonder (and the larger Wonder Full) is Meta's newest frontier model family, launched on October 12, 2026 for paid Muse plans, Muse Code and the Meta Model API. Muse Spark 1.3 powers the free tier, and Muse Spark Open is the open-weight model Meta released on October 15.
There is a free tier on Muse Spark 1.3. The $20 Power and $100 Maximum plans add Muse Wonder and more usage. Meta is rolling features out in waves, so check the Muse app for what you have.
No. This page covers Meta's Muse agent and models. The muse.ai video service and Microsoft's Muse game model are different products that share the name.
✨ Agent Stack Weekly · free
Get the week's best Muse tricks
What changed in Claude Code, Codex, Cursor, Copilot and Antigravity: limits, pricing, new MCP servers and the tools that actually work, with honest notes.