Tutorials

Begin Using Meta Muse in Python with Spark and Glimmer

Python guide for Meta Muse Spark API and open-weight Glimmer, with product distinctions, streaming, tools, typed output, and pricing.

Begin Using Meta Muse in Python with Spark and Glimmer

Key Takeaways

  • Meta’s “Muse” name covers multiple unrelated products: developer models, a consumer AI agent, hardware/gadgets, and older research or third-party products; only the Muse models are callable from Python via the Meta Model API or OpenRouter.
  • Muse Spark 1.3 is a hosted, closed-weight reasoning model that uses an OpenAI-compatible API, supporting streaming, image input, tool calls, and structured output with minimal code changes.
  • Muse Glimmer 30B is an open-weight model under Apache 2.0, not available through the Meta Model API, and can be self-hosted or used via hosted providers, often at lower cost than Spark.
  • The right entry point depends on the use case: Spark for OpenAI-style Python development, Glimmer for permissive open-weight hosting, and the Muse app for non-developer personal agent tasks such as email, shopping, and travel booking.

Developers who already use the `openai` package can reach Meta's Muse Spark model through the Meta Model API or OpenRouter with only a few lines of code, because Meta's API speaks the same protocol as OpenAI's. This tutorial is for Python developers who want to try the Muse family: you will call the hosted reasoning model Muse Spark, stream a conversation, send an image, run a tool call, and parse typed output, then compare it with the open-weights Muse Glimmer 30B. All timings and costs below come from the test runs described in the source material.

Begin Using Meta Muse in Python with Spark and Glimmer

One naming warning first: Meta uses "Muse" for three different things, and only one of them is callable from Python.

Muse Models, the Muse Agent, and Muse Charm Are Different Products

The Muse models are Meta's developer lineup, reached through the Meta Model API with an API key and usage-based billing. As of November 2026 the lineup is:

ProductAPI IDWeightsPrice
Muse Wonder 1.0 (preview)muse-wonder-1.0Closed$4.00 input and $20.00 output per 1M tokens
Muse Spark 1.3muse-spark-1.3Closed$1.25 input and $4.25 output per 1M tokens
Muse Spark OpenNot on the Model APIOpen, Muse Model LicenseFree to self-host on a multi-GPU server, or through hosted providers
Muse Ember 1.0muse-ember-1.0Closed$0.10 input and $0.50 output per 1M tokens for requests up to 200K tokens
Muse Glimmer 30BNot on the Model APIOpen, Apache 2.0Free to self-host, or from $0.30 input and $1.10 output per 1M tokens hosted
Muse CodeCommand-line tool `muse`ClosedAPI rates, or plans from $5 a month
Muse Imagemuse-image-1.0Closed$0.01 per image
Muse Voice Transcribemuse-voice-transcribe-1.0Closed$0.18 per hour of audio

Muse on its own is Meta's consumer personal AI agent, introduced September 8, 2026. It sends emails, books travel, and shops through connectors such as Instacart, and it is reached through iOS and Android apps, the web at muse.ai, WhatsApp, and a Mac desktop app. It is available only in the US for now, free up to a weekly usage limit, with Power and Maximum plans at $20 and $100 a month. It has no public API. Meta says the agent is powered by "Muse Spark, Meta's most capable model to date, built for real-world agentic work," without naming a version. To reach other people's Muse agents you submit an existing Model Context Protocol (MCP) server or an API to the Muse Connector Platform, which Meta reviews for "functional, security, and legal requirements." Meta's developer documentation does not tie a Power or Maximum plan to Model API usage, so budget for the app and the API separately.

Muse Charm is a gadget announced at Connect on September 23, 2026, with Meta promising only "more to share later this year." It has not shipped and has no price. Press reports describe a keychain-sized device with a two-inch touchscreen and 5G, but sources disagree about its camera, so treat those specs as provisional.

Begin Using Meta Muse in Python with Spark and Glimmer

There is also Muse Gadgets: on October 2, 2026, Meta open-sourced the Muse Gadget SDK under Apache 2.0, with firmware for ESP32 boards and a Linux SDK for machines such as a Raspberry Pi. The Python part is a package called `musegadget` that needs Python 3.9 or newer; it is not on PyPI, so you install it with the repository's `install.sh` script. After pairing a device with an SDK token from gadgets.muse.ai and enabling Developer mode in the Muse app, Muse can run commands on the machine, and scripts can post into your Muse chat with `musegadget send-user-msg`.

Older products called Muse, such as the InteraXon meditation headband, Microsoft's gameplay model, and Google's text-to-image research, are unrelated.

Which Entry Point Fits Your Situation

  • If you already use the `openai` package and want a hosted reasoning model with a 1M-token context, use Muse Spark through the Meta Model API or OpenRouter. OpenRouter charges the same token prices.
  • If you want open weights you can host on your own hardware under a permissive license, use Muse Glimmer with Ollama or a hosted provider.
  • If you want a personal assistant that sends emails and shops for you and you don't plan to write code, use the Muse app.

Calling Muse Spark From Python

Because the request format matches OpenAI's, the `openai` package works unchanged once you point it at Meta's endpoint and supply a Model API key. The steps:

Begin Using Meta Muse in Python with Spark and Glimmer
  1. Create a Model API team account, add a payment card, and generate an API key.
  2. Install the `openai` package and set the base URL to Meta's endpoint with your key.
  3. Send a chat completion request to `muse-spark-1.3`.
  4. Stream the response so tokens arrive as they are generated.
  5. Add an image to the message content to test vision.
  6. Define a tool and let the model request a call, then return the result.
  7. Ask for structured output and parse the typed result.

Reasoning is always on for Muse Spark. In the test run, lowering the effort setting to "minimal" cut latency by more than half and cost by about 60 percent without losing accuracy. That makes effort the first knob to tune when you are prototyping and watching spend.

Muse Spark handles images, tool calls, and structured output through the standard OpenAI request format, so existing code for those features carries over. Data from the standard tier is not used for training, according to the source material.

Running Muse Glimmer 30B

Muse Glimmer is a 30-billion-parameter model with downloadable weights under Apache 2.0, which allows commercial use. You can run it with Ollama or a hosted provider; hosted pricing starts at $0.30 input and $1.10 output per 1M tokens. In the test, hosted Glimmer answered the same test question for about one-eighth of Spark's cost. It is not on the Model API, so you will not reach it with a Meta API key; you point your client at whichever host you chose.

What This Means for Your Stack

The practical takeaway is that Meta's developer models slot into OpenAI-compatible tooling with almost no code change, which lowers the cost of trying a second provider. The trade-off is naming confusion: the consumer agent, the gadget, and the models share a brand but not an interface, and only the models are callable with an API key. Watch for Muse Charm's release details and any change to the preview pricing of Muse Wonder 1.0; until then, start with Muse Spark at minimal effort and benchmark Glimmer if self-hosting or permissive licensing matters to you.

Common Mistakes

  • Trying to call the Muse consumer agent with a Model API key; it has no public API.
  • Assuming a Power or Maximum app plan covers Model API usage; they are billed separately.
  • Leaving reasoning effort at its default while prototyping and paying for latency you don't need.
  • Looking for Muse Glimmer on the Model API; it is open weights only.
  • Expecting `musegadget` from PyPI; install it with the repository's `install.sh` script.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.