Muse MuseConcept
Muse MuseConcept

Independent concept built on Meta's Model API. Not an official Meta product.

Models

What Muse Muse builds with: Meta's hosted models, and the open model its lab tunes.

12 models shown

Meta's models

  • Muse Wonder

    New

    Context 1M · Chat

    muse-wonder-1.0
    Preview

    Every platform role: writing briefs and checking them, Muse Code's builds and fixes, and judging screenshots. The default for a product's own text and vision calls, and the fine-tuning lab's writer and judge.

    Reasoning is billed as output, so even short answers cost a few hundred tokens.

    Try in playgroundModel specsDocumentation
    Input
    $4 / 1M tokens
    Cached input
    $0.20 / 1M tokens
    Output
    $20 / 1M tokens
    Calls, 30 days
    780

  • Muse Spark Open

    Context 1M · Open weights

    Open-weight Muse Spark: 816B total parameters, 42B active, mixture of experts; text, images, audio and video in, text out; self-host with SGLang or vLLM.

    DownloadModel specsDocumentation
    Access
    Download and self-host
    License
    Muse Model License

  • SAM 3.1

    Image and video segmentation

    sam-3.1

    Inside products, to find, outline and track objects in photos and video.

    Try in playgroundModel specsDocumentation
    Image
    $0.0025 per image
    Video
    $0.0002 per frame, 50 frames at least
    Calls, 30 days
    145

  • Muse Spark 1.3

    Context 1M · Chat

    muse-spark-1.3

    Meta's model for agentic and coding work. Muse Muse doesn't call it: anything that would run on it runs on Muse Spark 1.3 Contributor.

    Model specsDocumentation
    Input
    $1.25 / 1M tokens
    Cached input
    $0.15 / 1M tokens
    Output
    $4.25 / 1M tokens

  • Muse Spark 1.3 Contributor

    Context 1M · Chat

    muse-spark-1.3-contributor

    The model Muse Muse builds on right now: the brief, the build, the checks and a build's own test calls. Meta may train on what it's sent, so don't put personal, sensitive or confidential information in a build. Live products don't run on it.

    Older products built on Muse Spark 1.3 run on it instead, since Muse Muse doesn't call Spark 1.3 itself.

    Try in playgroundModel specsDocumentation
    Input
    $0.10 / 1M tokens
    Cached input
    $0.002 / 1M tokens
    Output
    $0.20 / 1M tokens
    Calls, 30 days
    3,805

  • Muse Voice Transcribe

    Speech to text

    muse-voice-transcribe-1.0

    Turns a voice memo into text before the brief, and transcribes recordings inside products.

    Try in playgroundModel specsDocumentation
    Audio
    $0.18 per audio hour
    Calls, 30 days
    134

  • Muse Glimmer 30B

    Context 128K · Open weights

    Meta's open model: 30B parameters, reads text and images. Its license is Apache 2.0, with Meta's usage policy. It isn't on Meta's Model API, so Muse Muse's lab runs it on rented GPUs to tune it and to run live demos. A product can pick it for local or offline use, through a server someone hosts.

    New fine-tunes start from Muse Muse's 2026-knowledge Glimmer, OfflineBase/Glimmer-30B-2026-Knowledge (opens in a new tab), and so do the live demos. Muse Spark Open wrote part of its training data, so it and every model tuned from it are Built with Muse Spark Open, under the Muse Model License. For people 18 and over only. Its answers are written by a model. Never present them as written by a person. It can make up answers, and it can go along with a false premise in a question. Check what it says before you rely on it.

    DownloadModel specsGlimmer docsLab results
    Access
    Download and self-host
    License
    Apache 2.0

Muse Muse and the products it ships call the hosted models through Meta's Model API. Prices are Muse Muse's estimates. Calls are from Muse Muse's usage records for the last 30 days.

Tuned Glimmer models

Where they run

  1. Meta's Model API

    The hosted models run on Meta's servers. Muse Muse calls them over the API.

  2. A rented GPU

    The lab tunes Glimmer and runs its live demos on Lambda GPUs, shut down after each job.

  3. On a device

    A tuned model with public GGUF files runs on your own computer, offline.

What fine-tuning did on open weights: the method, the cost, the wins, the losses and every answer.