Muse is Meta's personal AI agent, launched September 8, 2026 with the tagline "AI that gets it done for you." Unlike a chatbot that answers questions, Muse is built to do things: sort email, book reservations, shop, manage calendars, fill forms, build fitness plans – running on a dedicated Secure VM with its own browser, filesystem and terminal. It is powered by Muse Spark 1.3, the flagship of Meta Superintelligence Labs' fast-growing model family.
The catch, for most of this site's readers: Muse is currently US-only, 18+, and the models that matter most for local AI are a separate story – including a 30B open-weights model you can already run on your own hardware. This guide covers the agent, the pricing, the full Spark model family, the developer API, and what it all means for local-AI users.
Source note: facts below come from Meta's announcements, developer docs, and launch coverage (AP, PBS, TechCrunch, WIRED, CNBC). Benchmark figures are vendor-reported unless stated otherwise.
The short version: Muse in 60 seconds
- What it is: a personal agent you message – it picks the skills a task needs, browses the web, pulls from connected apps, and acts across multiple steps, with permissions and approvals you control.
- Launched: Sept 8, 2026. US only, 18+. Web at
muse.ai, dedicated iOS/Android Muse app, WhatsApp chat, Mac app; Meta AI glasses support coming. - Price: free with a usage limit; paid subscriptions at $20/mo and $100/mo reported. A payment card is required to start, since billing kicks in as usage grows.
- Under the hood: Muse Spark 1.3 (1M-token context) running in a per-user Secure VM – full browser, filesystem, terminal – isolated from Meta's other systems.
- Developers: Meta Model API serves Spark 1.1/1.2/1.3 at $1.25 input / $4.25 output per million tokens ($0.15 cached), with a Contributor tier at $0.10/$0.20 and $20 in free credits.
- Local AI angle: Muse Glimmer is a 30B Apache-2.0 open-weights agent model (Aug 10) that runs on a Mac or single consumer GPU – and Meta has promised Spark open weights too.
What Muse actually does
Message it like a capable assistant: sell your car, book a plane ticket, sort the inbox, build a fitness plan, turn a long-term goal into an action plan. Because it's agentic, it figures out which built-in skills a job needs and gets to work – browsing, pulling information from connected apps, and chaining steps without being micromanaged.
Connections are the interesting part. Muse ships with built-in connectors for common services, and where a service offers a public API it can wire up a connection from credentials you provide. Where no API exists, it falls back to operating the service through its own browser – the classic agent escape hatch, running inside your Secure VM rather than on your laptop.
The Secure VM: Meta's trust pitch
Every outlet covering the launch – AP, PBS, TechCrunch, WIRED – led with the same question: will consumers trust a Meta agent with their email, calendar and payments? Meta's answer is architectural: Muse runs in a dedicated, secure virtual machine that houses both the agent and your data, with automated protections, agentic-AI security standards, per-action permissions and review/approve checkpoints, plus a protective layer meant to stop your information being shared without consent. Meta published a full "How We Built Safety Into Muse" account on Sept 8, 2026.
Honest framing: a VM boundary plus approvals is genuinely better than a chatbot with your password in its context window – but it is still a cloud computer acting on your accounts, from a company whose business is advertising. If your threat model excludes US clouds entirely, Muse is not for you no matter how good the sandbox is. That is exactly the gap local agents fill – more below.
Where to get Muse, and what it costs
| Access point | Status |
|---|---|
Web at muse.ai | Live (US) |
| Muse app (iOS / Android) | Live (US), works even with the app closed |
| WhatsApp chat | Live (US) |
| Muse for Mac | Live – organizes files, fills forms, pulls from Messages/Calendar/Notes with permission |
| Meta AI glasses | Coming soon |
| Tier | Price | Notes |
|---|---|---|
| Free | $0 | Usage limit; refreshes |
| Paid (reported) | $20/mo | Higher usage; card required at signup |
| Paid top (reported) | $100/mo | Heaviest use |
Availability is the headline limitation: United States only, 18 and over, with no announced dates for India, Europe or elsewhere. Readers outside the US can use the underlying models through the developer API today – or run the open-weights sibling locally right now.
Under the hood: the Muse Spark family
| Model | Released | What it is |
|---|---|---|
| Muse Spark | Apr 8, 2026 | First MSL model: natively multimodal reasoning, tool-use, visual chain-of-thought, multi-agent orchestration; Contemplating mode (parallel agents) hit 58% on Humanity's Last Exam |
| Muse Spark 1.1 | Jul 9, 2026 | Agentic upgrade (tool/computer use, coding); launched the public Meta Model API preview; 1M-token context |
| Muse Spark 1.2 | Aug 5, 2026 | Coding-focused; arrived with Muse Code, a terminal coding agent for macOS/Linux |
| Muse Spark 1.3 | Sep 2, 2026 | Current flagship powering Muse: long-horizon agentic + coding, ~20% fewer tool calls and ~25% fewer tokens than 1.2; reasoning modes minimal → xhigh, with max reasoning following safety testing |
| Muse Glimmer 30B | Aug 10, 2026 | Open-weights (Apache 2.0) agent model for local hardware – see below |
Spark 1.3 is where Meta claims to have joined the frontier conversation: Artificial Analysis briefly charted 1.3-xhigh at the intelligence-vs-cost frontier (~$0.55 per task), edging past Gemini 3.8 Flash within hours of Google's own lead. Meta's own evaluation methodology compares 1.3 at max effort against 1.2, Claude Opus 5 and GPT-5.6 Sol. Treat both as directional – vendor and third-party charts, not controlled head-to-heads.
Developer API: prices and practical details
| Per 1M tokens | Standard | Contributor |
|---|---|---|
| Input | $1.25 | $0.10 |
| Cached input | $0.15 | $0.002 |
| Output | $4.25 | $0.20 |
The Standard tier (1.3, 1.2, 1.1) never trains on your data; Contributor (1.3/1.2) trades steep
discounts for permission to train on your prompts and completions – built for prototyping. Context is
1,048,576 tokens on all Spark versions, the endpoint is OpenAI-SDK compatible
(https://api.meta.ai/v1), every account starts with $20 in free credits, and the models
are also on OpenRouter and LiteLLM (day-0 support for 1.3). Sibling endpoints: Muse Voice Transcribe
($0.18/hr) and Muse Image ($0.01/image, with Meta's invisible Content Seal watermark); Muse Video is
in preview. One caveat from Meta's own docs: audio understanding in 1.3 is currently degraded – use
1.2 or Voice Transcribe for audio work.
If you estimate agent bills, run the numbers in our AI API cost calculator – at $1.25/$4.25, Spark undercuts Opus-class $5–$10/$20–$50 pricing substantially, which is exactly why it briefly topped cost-efficiency charts.
Muse Glimmer: the one you can run yourself
This is the part local-AI readers should care about most. Muse Glimmer (Aug 10, 2026) is a ~30B dense multimodal model (28B decoder + ~2B perception encoder), distilled from Muse Spark and released under Apache 2.0 on Hugging Face – Meta's first open model from Superintelligence Labs, with Spark 1.2 weights promised to follow. It is explicitly trained for the tool-use loop: multi-step plans, sequential tool calls, failure recovery, runtime memory, vision input – and its model card names OpenClaw and Hermes Agent as tested scaffolds, with llama.cpp, vLLM, Ollama, LM Studio, Jan, MLX and ExecuTorch support.
| Benchmark (vendor-reported) | Glimmer 30B | Gemma4-31B | Qwen3.6-27B |
|---|---|---|---|
| MCP Atlas (agentic) | 75.5 | 54.2 | 62.5 |
| DeepSearch QA | 74.6 | 61.7 | 71.1 |
| SWE-Bench Verified | 76.0 | 66.6 | 77.2 |
| SWE-Bench Pro | 51.2 | 36.9 | 50.2 |
| AIME 2026 / GPQA Diamond | 94.7 / 83.5 | – | – |
| GDPval-AA v2 | 953 | 811 | 1141 |
| OSWorld-Verified | 65.9 | 58.5 | 75.6 |
Read it honestly: Glimmer leads most agentic rows but loses knowledge-work (GDPval) and computer-use (OSWorld) to Qwen – and every figure is self-reported, without independent replication yet. Still, a 30B Apache-2.0 agent model beating its weight class on tool-use benchmarks is exactly what the local ecosystem needed.
Running Glimmer on your hardware
Quant ~memory Example hardware ───────────────────────────────────────────────── 4-bit (UD) 17GB+ 24GB GPU / 32GB Mac 6-bit 20–22GB+ RTX 5090 / 48GB Mac 8-bit 34GB+ 128GB Mac / DGX Spark BF16 58GB+ 128GB Mac / DGX Spark
Meta's 4-bit build squeezes the model under 20GB, leaving room for KV cache, the perception encoder and a speculative-decoding drafter – Meta's DFlash drafter reportedly pushes an RTX 5090 from ~75 to ~233 tok/s (3.1×). Tested scaffolds include OpenClaw and Hermes; recommended settings are temperature 1.0, top_p 0.95, top_k 64, with controllable reasoning (low → xhigh) and up to 131K–262K context. Check fit first with our LLM VRAM calculator and VRAM-per-model guide, and see our Apple Silicon guide or best GPU for local LLMs for the machine side.
Muse vs the alternatives
WIRED frames the fight as Muse vs OpenClaw and Instinct – the always-on, user-controlled agent crowd. The honest split: Muse wins on zero-setup convenience (message it in WhatsApp, done) inside the US; OpenClaw-style local stacks win on privacy, offline use and no per-token meter – and Glimmer just made those stacks much stronger. Against Claude and ChatGPT agents, Muse's differentiators are the Secure VM story and Meta-ecosystem reach (WhatsApp, Instagram, glasses); its weaknesses are US-only availability and a thinner connector catalogue that will live or die on how fast Meta ships integrations.
What it means for local-AI users
Two things. First, Glimmer is a genuinely good day for consumer hardware: an Apache-2.0 30B agent brain that fits a 24GB card or a 32GB Mac, speaks OpenClaw/Hermes natively, and costs nothing per token. If you run always-on agents, benchmark it against your current Qwen/Gemma stack – see our local vs cloud AI tradeoff guide for the math. Second, Meta's promised Spark open-weights release (Zuckerberg: 1.2 "soon") is the one to watch – a frontier-class open release would cascade down the way open weights always do. Until then, cloud Muse stays a US-only curiosity for most readers, while Glimmer is downloadable today.
Bottom line
Muse is Meta's most serious agent play yet – a real Secure-VM computer of your own in the cloud, on WhatsApp and in your pocket, at prices ($20/$100) that signal who it's for. But the trust ask is enormous, availability is one country, and the most interesting artifact of the whole launch might be the free one: Glimmer, the 30B open model that lets you build a Muse-like agent that never phones home. For this site's readers, that ordering is exactly right – cloud convenience vs local control, and for once you can try both sides.
Sources
- Meta: Muse – personal AI agent (features & capabilities)
- muse.ai – official product site
- Meta Research: How We Built Safety Into Muse (Sep 8, 2026)
- Meta Research: Introducing Muse Spark 1.3 (Sep 2, 2026)
- Meta: Introducing Muse Spark (Apr 8, 2026)
- Meta Developers: build with Muse Glimmer (Aug 12, 2026)
- Meta Model API docs: pricing and rate limits
- Meta Model API docs: models
- TechCrunch: Meta debuts its Muse AI agent (Sep 8, 2026)
- WIRED: Muse needs you to trust it (Sep 8, 2026)
- CNBC: Meta launches Muse Glimmer (Aug 10, 2026)
FAQ
What is Meta Muse?
Meta's personal AI agent, launched Sept 8, 2026: message it to get tasks done (email, shopping, bookings, calendars) via a Secure VM with its own browser, filesystem and terminal, powered by Muse Spark 1.3.
How much does Muse cost?
Free with a usage limit; paid tiers around $20/mo and $100/mo reported for heavier use. A payment card is required at signup since billing scales with usage.
Where is Muse available?
US only, 18+. Web at muse.ai, iOS/Android Muse app, WhatsApp chat and a Mac app; Meta AI glasses support is coming. No dates announced for other countries.
What models power Muse?
The agent runs on Muse Spark 1.3 (1M context). Developers get Spark 1.1/1.2/1.3 via the Meta Model API, and Muse Glimmer 30B is the Apache-2.0 open-weights sibling for local hardware.
How much does the Muse Spark API cost?
Standard: $1.25 input / $4.25 output / $0.15 cached per million tokens; Contributor (trains on your data): $0.10 / $0.20 / $0.002. New accounts get $20 in free credits.
Can I run Muse models locally?
Muse Glimmer 30B, yes – ~17GB+ at 4-bit via Ollama, LM Studio, llama.cpp, vLLM or MLX on a 24GB GPU or 32GB Mac. Spark itself is API-only, with an open-weights release promised.