Meta Muse: Every Model in Meta's Personal Agent Family, and Which One You Actually Want

There is no single thing called Meta Muse. Depending on which link someone sent you, it means the consumer agent Meta put into US circulation on September 8, 2026 — or one of the models underneath it: Muse Spark, the natively multimodal reasoning model that does the actual thinking; Muse Image and Muse Video, the media generators Meta previewed on July 7, 2026; Muse Glimmer, the 30-billion-parameter open-weight model under Apache 2.0; Muse Code, the terminal coding agent; Muse Voice Transcribe, the speech-to-text model; and Muse Realtime Avatar, the embodiment layer Meta showed at Connect 2026 on September 23. Two of those names are also things that are not models at all: Muse Connectors is a product surface, and Muse Realtime Voice is the conversational layer inside the agent.

This page exists to sort that out. It is not a news article about a dated event. It is the canonical reference page for a model family that is live in our catalogue today — the page a reader reaches after searching the model's own name. Its warrant is standing search demand that we measured first-party on 2026-09-29, not the freshness of a launch. The datable anchors in it (September 8 for the agent, September 23 for Connect) are facts used to date these products, not events this page is reporting. We covered the agent launch separately and this page does not re-report it; if you want that story, the launch coverage is the place to go.

The one-minute routing table, in prose

Start with the question you are actually asking, and stop at the line that answers it.

• If you want the consumer product — a personal agent that reads your email, books things, runs in the background and asks permission before it spends money — that is Meta Muse, the agent, and its surface is the Muse app or the web at muse.ai. Meta's announcement post lists US availability on iOS, Android and muse.ai, and describes it as free "for most of what people need," with subscription plans above that. That framing is Meta's own, published on Meta's own page, and it is the only place a free tier is claimed; nothing about rate limits or credit allowances is published.

• If you want the model that answers questions and writes code — the one you would call from an API — that is Muse Spark, and the current identifier is muse-spark-1.3. Meta's own model page documents a 1M-token context window at $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens, with a cheaper contributor tier at $0.10 / $0.002 / $0.20 that opts your traffic into product improvement.

• If you want to generate or edit a still image — that is Muse Image, an agentic image model that Meta prices at $0.01 per image on its own model page. It is not a text-in, text-out model; it takes a prompt and reference images and returns pictures.

• If you want generated video — that is Muse Video, and this is the one member of the family where almost everything is still unannounced. Meta previewed it on July 7, 2026, and has published no release date, no pricing, no clip length, no resolution and no regions since. Anything quoting a specific launch day or a price for Meta Muse Video without a newer Meta announcement is speculating, and that includes us.

• If you want to transcribe audio — that is Muse Voice Transcribe, priced by Meta at $0.18 per hour ($3.00 per 1,000 minutes) on its model page, with speaker attribution. It is not a text-in model; it consumes audio and emits text.

• If you want to run an agent on your own hardware with no API call at all — that is Muse Glimmer, 30 billion parameters, Apache 2.0, small enough for a single consumer GPU or a Mac, and distributed by Meta as downloadable weights rather than through an endpoint.

• If you want an agent in your terminal or your CI — that is Muse Code, a CLI that drives Muse Spark, installed with a one-line shell script, and Meta now documents both a macOS/Linux installer and a Windows PowerShell installer.

• If you want a face and a voice on the other end of a conversation — that is Muse Realtime Avatar, announced September 23, 2026 alongside the research post Meta titled "Bringing Your Muse to Life."

What Meta's own chronology says

The family is easier to hold in your head as a sequence than as a list, because Meta has shipped it in a steady cadence through 2026. These dates come from Meta's research blog index, which we read directly rather than from anyone's summary of it:

• April 8, 2026 — "Introducing Muse Spark: Scaling Towards Personal Superintelligence," the first Muse model out of Meta Superintelligence Labs.

• July 7, 2026 — Muse Image and Muse Video announced together. Image shipped; Video did not.

• July 9, 2026 — Muse Spark 1.1.

• August 5, 2026 — Muse Code and Muse Spark 1.2 on the same day.

• August 10, 2026 — Muse Glimmer, the open-weight release.

• September 1, 2026 — Muse Voice Transcribe.

• September 2, 2026 — Muse Spark 1.3.

• September 8, 2026 — the Muse agent itself, plus Meta's companion post "How We Built Safety Into Muse."

• September 23, 2026 — "Bringing Your Muse to Life," the realtime avatar work, published on the opening day of Connect.

One correction to make here, because this family is unusually easy to get wrong on dates and one of the dates in wide circulation is wrong. Meta's Connect recap post is dated September 24, 2026 and opens "Yesterday at Connect 2026." It says Meta is bringing Muse to its AI glasses "in the coming months." Glasses access was announced on September 23; it is not shipping, and Meta's own page does not describe it as available. If you have read that Muse reached AI glasses on September 23, that is an announcement being reported as an availability, and Meta's page wins.

The developer surface, and why it matters more than the app

The part of this family a builder should care about is not the agent. It is Meta Model API, because it is the first time Meta has exposed a Muse model through a surface that looks like something you already own. Meta's own quickstart says it plainly: Model API "is drop-in compatible with the OpenAI SDK, the Anthropic SDK, and OpenAI-compatible agent CLIs, so most stacks work with a base-URL and key change." The base URL is https://api.meta.ai/v1 and the model ID for the current Spark is muse-spark-1.3.

Whether that compatibility is real in practice is a separate question from whether it is true on paper, and we have not run it. What is verifiable today is that Meta documents three protocol paths rather than one — a Responses API, a Chat Completions API and a Messages API — which is a stronger signal than a single "OpenAI-compatible" endpoint usually is, because it means the Anthropic-format path was built deliberately rather than bolted on.

Modalities are the place to read carefully. Meta's documented Muse Spark input modalities are text, image, PDF and video, with text output. That is a natively multimodal model with tool use, not a text model with an image endpoint next to it. If you need output that is not text, Spark is not the member of the family you want — that is the whole reason Muse Image and Muse Voice Transcribe exist as separate models with separate prices.

This is also where routing earns its place. A family this wide means the model you want depends on the request, and the cost gap between members is large: $0.01 per image and $0.18 per audio hour sit in a completely different budget envelope from $1.25 per million tokens. Putting several of them behind one endpoint, with failover when a provider degrades and a routing rule that picks the model per request type, is a smaller engineering problem than wiring each vendor's SDK separately — and because a pass-through router bills provider list price, a Meta price change lands the same day rather than at the next contract renewal.

Muse Glimmer is the one you run yourself

Glimmer is worth separating out from the rest, because it inverts the distribution model. Meta calls it "an open model built for always-on local agents," at 30 billion parameters under Apache 2.0, and you get it by downloading the weights rather than by calling an endpoint. Meta's own positioning is that it is small enough for a single consumer GPU or a Mac, and it names tool calling, persistent state across restarts and self-managed memory across long sessions as the design targets.

Meta publishes a comparison table on Glimmer's model page against Gemma4-31B and Qwen3.6-27B in thinking mode. It is a vendor-selected comparison on vendor-chosen benchmarks, and it should be read that way — Glimmer leads on most of the agentic rows Meta chose to print, and trails on GPQA Diamond and Humanity's Last Exam. We are not repeating the individual figures here because they belong to a model we do not serve; if you are choosing a local model, read the numbers off Meta's own model page and its published model card directly and run your own task against it. Note also that an earlier version of our own coverage described Muse Code as macOS and Linux only. Meta's current documentation lists a Windows PowerShell installer as well, so that claim is stale and Meta's page wins.

The security architecture, in one paragraph

Muse's security design is the most substantive engineering Meta has published about the family, and it is worth knowing in outline before you decide how much access to give a personal agent. The agent runs inside Muse Secure VM — a dedicated Linux VM per user, with the agentic harness inside a systemd-nspawn container whose root is mapped to an unprivileged host user, so container root is not host root. Credentials never reach the agent: they live in storage owned by hatch-authd, which mints surrogate tokens, and a separate Sentinel agent is the sole permission authority for connector actions and for all network egress. Sentinel evaluates a request at both L4 and L7, does just-in-time credential substitution at the network boundary, and uses kernel-level data-flow tracking (Meta calls it "tainted egress" — implemented with eBPF) so that a process that has read your data cannot ride an auto-allow path. Safety classifiers run outside the runtime cell precisely so that an attacker cannot switch them off from inside it. Meta's own numbers for the bounty program are the useful calibration: up to $300,000 for valid reports, including up to $130,000 for a prompt injection that affects a single user. That is a company publishing a price on the failure mode of its own product, which is not the norm. The full write-up is at research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse and it is worth twenty minutes.

What Meta has not published

A reference page that only repeats marketing is not a reference page. Here is the honest state of the gaps:

• Monolithic benchmarks. There is no single published index number that ranks this family against frontier peers on Meta's own pages. Muse Spark 1.3's page prints a small table comparing GDPval-AA v2, JobBench and one partial column against Muse Spark 1.2, GPT 5.6 Sol and Opus 5, and the values are Meta-measured at Meta-selected effort settings. That is an internal ablation, not a head-to-head, and it does not extend to the agent, Image, Video, Glimmer or Voice Transcribe.

• Any head-to-head at matched effort and harness. Two of these models have not been benchmarked against each other anywhere we could verify, in either direction, on the same evaluation rig. Where a number exists it is vendor-reported and we say so in the sentence carrying it.

• Muse Video's entire commercial shape. No date, price, cap or region, as covered above.

• Rate limits and credit structures. Meta's pricing and rate-limit documentation page returns a shell with no readable figures from here — a 107,8xx-byte document with no dollar values in it — so any number about Meta's caps would be invented. The one exception is the per-unit prices quoted above, which we read on Meta's individual model pages today.

• Anything about what ai.meta.com/muse/ says. That URL returns 200 with a confirming title and no body we can extract, so we are not going to characterise its contents.

What is actually callable, and where

This is the part where availability claims go stale fastest, so it is stated narrowly and from a live check rather than from a sibling article. As of 2026-09-29 the OrcaRouter catalogue serves exactly two Muse routes, verified individually against the public model API:

• Meta: Muse Spark 1.2 — 1,048,576-token context, $1.25 per million input tokens, $0.15 per million cached input tokens, $4.25 per million output tokens, listed release date 2026-08-05.

• Meta: Muse Spark 1.1 — same 1,048,576-token context and the same $1.25 / $0.15 / $4.25 rate, listed release date 2026-07-16.

Those two prices are the same as the rate Meta publishes for muse-spark-1.3 on its own model page, which is what a zero-markup pass-through looks like: you pay Meta's list price, and we do not add to it. Muse Spark 1.2 is live on OrcaRouter at Meta's exact list price and is the one to reach for if you are porting an existing Spark workload.

Nothing else in the family is on our catalogue, and the absence is worth being blunt about: Muse Spark 1.3, Muse Image, Muse Video, Muse Glimmer, Muse Code and Muse Voice Transcribe all return a not-found from the public model API today. If you want the 1.3 identifier or the $0.01 image price, that is Meta's own endpoint, not ours. A confident wrong availability claim is worse than no claim, so that is the whole paragraph.

How to think about the family in one pass

Three separations do most of the work. First, product versus model: Meta Muse is a consumer agent with a UI and a permission system, and it is not something you call from code — the callable thing underneath it is Muse Spark. Second, endpoint versus download: every member except Muse Glimmer is reached through Meta Model API or a client built on it, and Glimmer is reached by downloading weights. Third, shipped versus announced: Image, Voice Transcribe, Spark and the agent are shipping; Video and glasses access are not, and both have acquired a false date in circulation.

If you are deciding what to build on this quarter, the practical read is that Muse Spark is the only member of the family with a published price, a published context window, a documented API and a cheap failover path — which makes it the low-risk entry point and makes everything else a thing to watch rather than a thing to plan around. If you are deciding what to read, the security post is the highest-value document Meta has published about the family, because it is the one that tells you what happens when the agent is wrong.