· 11 min read · Nadeera Silva
Meta Muse Models Explained: Wonder, Spark, Ember and Glimmer
Meta Muse now covers Wonder, Spark 1.3, Ember, Glimmer and a device called Charm. Here is what each one is, what it costs and which to use.

Meta Muse is no longer one model. In the last three months Meta has shipped Muse Wonder, Muse Spark 1.3, Muse Ember, an open-weight Muse Spark, a small local model called Muse Glimmer, and announced a gadget called Muse Charm. If you build with LLMs, the names alone make it hard to know what you can call from code, what you can download, and what is just a product. This guide maps the whole Meta Muse lineup from Meta's own docs and model cards, shows what each one costs, and gives a small router you can use to put the right model on each part of an agent.
What is Meta Muse?
Muse is Meta's brand for its current generation of AI: the Muse assistant app, the Muse agent, the Muse Code terminal coding agent, and the models behind them. For developers, the important split is between three kinds of thing that all carry the Muse name:
- Hosted reasoning models you call through the Meta Model API: Muse Wonder, Muse Spark and Muse Ember.
- Open-weight models you download and run yourself: Muse Spark Open and Muse Glimmer.
- Products built on top of the models: the Muse app, Muse Code, and the Muse Charm device.

The Meta Muse models at a glance
| Name | What it is | How you get it | Price (per 1M tokens, in / out) | Released |
|---|---|---|---|---|
| Muse Wonder | Frontier reasoning model for hard coding and long agent runs | Model API (muse-wonder-1.0), Muse Code | $4.00 / $20.00 | 12 Oct 2026 |
| Muse Wonder Full | Higher-capability Wonder variant | Muse Code Power ($50/month) and Maximum ($200/month) plans only | Subscription | 12 Oct 2026 |
| Muse Spark 1.3 | Mid-tier model for the main agent loop | Model API (muse-spark-1.3), Muse Code | $1.25 / $4.25 | 2 Sep 2026 |
| Muse Ember | Smallest hosted model for high-volume steps | Model API (muse-ember-1.0), Muse Code | $0.10 / $0.50 | 9 Nov 2026 |
| Muse Spark Open | Open-weight Muse Spark, 816B total / 42B active | Hugging Face, Muse Model License | Your hardware | 15 Oct 2026 |
| Muse Glimmer 30B | Small open model for local agents | Hugging Face, Ollama, Apache 2.0 | Your hardware | Model card dated Aug 2026 |
| Muse Charm | Pocket device for talking to the Muse agent | Not on sale yet | No price announced | Announced 23 Sep 2026 |
Prices are Standard mode list prices from Meta's pricing page. Muse Ember charges more for prompts over 200K input tokens ($0.30 in, $0.012 cached, $1.50 out), while Wonder and Spark charge the same rate at any context length.
Muse Wonder and Muse Wonder Full
Muse Wonder launched on 12 October 2026 as Meta's top model on the API. Meta positions it for multi-file changes in real repositories, multi-hour agent runs, sub-agent orchestration and computer use. It costs $4 per million input tokens, $0.20 cached and $20 output.
Muse Wonder Full is the confusing one. According to Meta's launch post, it is available only inside Muse Code on the Power Usage ($50 a month) and Maximum Usage ($200 a month) plans. The Choose a model guide is explicit that no model ID on the Model API selects it. So if a tutorial tells you to set model="muse-wonder-full" in your API code, it is wrong.
Muse Spark 1.3 (and 1.2, 1.1)
Muse Spark 1.3 came out on 2 September 2026 and is the model Meta recommends for most work. Meta says it uses about 20% fewer tool calls and 25% fewer tokens than 1.2 in its internal comparisons. The older muse-spark-1.2 and muse-spark-1.1 IDs are still listed at the same price, which is why you still see searches like "meta muse spark 1.1". Unless you have a pinned evaluation on 1.1, there is no price reason to stay on it.
Spark is also the only hosted model with Contributor mode: muse-spark-1.3-contributor and muse-spark-1.2-contributor cost $0.10 in and $0.20 out, but Meta may use your prompts, files and outputs to train future models, and the rate limit drops from 3,000 to 150 requests per minute.
Two smaller details from the docs: the "max" reasoning effort works on muse-spark-1.3 but not on 1.2, 1.1 or the Contributor variant, and Meta notes that audio understanding may be degraded on 1.3, so it points audio work to 1.2 or Muse Voice Transcribe.
If you want the open-weight Spark rather than the API, my Muse Spark Open developer guide covers the model card, hardware and license in detail.
Muse Ember: the cheap workhorse
Muse Ember, released on 9 November 2026, is Meta's smallest hosted model. Meta describes it as built for "the work agents repeat all day": tool calls, coding subtasks, routing, extraction and long-document parsing. It has a 1M-token context window, up to 512K output tokens, and supports the Responses API.
The headline claim is cost. On Artificial Analysis Intelligence Index v5, Meta reports Ember at about $0.14 per task with a score of 36, against $2.39 per task for Muse Spark 1.3 at 37. Meta also says Ember matches Claude Haiku 5.5's best score at about a third of the cost and beats every GPT-6 Luna setting. Some partner numbers are striking too, for example Cloudflare reported up to 30% faster tail latency than the comparable model it tested.
Meta's own docs also say where not to use it: for hard single-shot coding problems, knowledge-heavy questions or visual coding, use Spark or Wonder instead. That matches what I would expect from a small model, and it is the right way to think about Ember: as the model for sub-agents and bulk jobs, not the brain of the agent.
Muse Glimmer: the local model
Muse Glimmer 30B is a different kind of release. It is an open model under Apache 2.0, distilled from Muse Spark and designed to run agents locally without a network connection. From the model card:
- About 29.6B parameters, dense transformer, plus a roughly 1.8B-parameter vision encoder. It takes text and images in and returns text. Audio is not supported.
- Context of 131,072 tokens or more, and a knowledge cutoff of 4 January 2026.
- Quantized builds target 24 GB or 32 GB of GPU memory. The Ollama build is an 18 GB Q4_K_M download.
- A speculative decoding drafter (DFlash) that Meta measured at 233 tokens per second on an RTX 5090, and about 50 on an M5 Max.
- Reasoning strength is set in the system prompt (
Reasoning strength: high), with low, medium, high and xhigh levels.
Meta's research paper was published on 19 October 2026, while the Hugging Face card lists an August 2026 release date, which explains why you see both dates online. Running it is one command:
ollama run muse-glimmer:30b
# Apple Silicon build with the MLX engine
ollama run muse-glimmer:30b-mlxGlimmer is the option to look at when data must not leave a machine, for example a support assistant that reads internal documents on an office workstation. Meta's card recommends adding human confirmation for irreversible actions when the model can act in the real world, which is good advice for any local agent.
Muse Charm is a device, not a model
Muse Charm shows up in the same searches, but it is hardware. Meta revealed it at Connect 2026 on 23 September as a small pocket device for talking to the Muse agent. Meta's 5 November post says voice conversations are "coming soon" to Muse Charm. There is no Muse Charm model ID, no API and, as of 12 November, no published price or ship date. You do not need to plan any integration work for it.
Which Muse model should you use?
Meta's Choose a model guide gives a clear workload table. Here is my condensed version for typical client projects:
| Workload | Model | Why |
|---|---|---|
| Coding agent main loop, multi-file refactors | Muse Wonder | Capability matters more than token price |
| Chatbot or assistant main loop on a budget | Muse Spark 1.3 | Strong reasoning at about a third of Wonder's input price |
| Tool calls, routing, classification, extraction | Muse Ember | Cheapest hosted option, built for many short steps |
| Same task over thousands of files or records | Muse Ember | Watch the 200K-token long-context price step |
| Data must stay on your own machine | Muse Glimmer 30B | Apache 2.0, runs on one 24 GB GPU |
| Self-hosting a frontier-size model | Muse Spark Open | Open weights, but server-class hardware |
| Audio understanding | Muse Spark 1.2 or Muse Voice Transcribe | Meta flags degraded audio on Spark 1.3 |
The pattern Meta recommends is to combine them in one agent: a bigger model plans on the main loop and Ember handles the volume. I use the same split for AI chatbots that need to look things up, where most turns are simple tool calls and only a few need deep reasoning. If you are weighing Muse against other providers, my LLM API pricing comparison shows how to compare cost per task instead of cost per token.
Calling the Muse models from code
The Model API is drop-in compatible with the OpenAI SDK at https://api.meta.ai/v1, per Meta's SDK docs. Keys start with LLM_, and the SDK does not read MODEL_API_KEY on its own, so pass it explicitly. Because Wonder, Spark and Ember accept the same request body, routing is just a different model value per call:
import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.meta.ai/v1",
api_key=os.environ["MODEL_API_KEY"], # keys start with LLM_
)
# One place to decide which Muse model handles which kind of step.
MODEL_FOR = {
"plan": "muse-spark-1.3", # main loop; use muse-wonder-1.0 for hard repo work
"tool_step": "muse-ember-1.0", # short, bounded sub-tasks
"extract": "muse-ember-1.0",
}
def run(step: str, prompt: str) -> str:
response = client.responses.create(model=MODEL_FOR[step], input=prompt)
return response.output_text
plan = run("plan", "Break this support ticket into lookup steps: order #1234 never arrived.")
print(plan)
print(run("extract", "Return only the order number from: 'order #1234 never arrived'"))Keep the mapping in config, not scattered through the code, so you can move a step to a bigger model when your evaluations say it needs one.
What a task costs on each model
A quick worked example with list prices, for a step with 20,000 input tokens and 2,000 output tokens:
| Model | Cost per step | Cost per 1,000 steps |
|---|---|---|
| Muse Wonder | $0.120 | $120.00 |
| Muse Spark 1.3 | $0.034 | $33.50 |
| Muse Ember | $0.003 | $3.00 |
Two things change this in practice. Reasoning tokens bill as output, as LiteLLM's Spark 1.3 notes point out, and Wonder and Ember default to "max" effort when a request sets none, so real output counts can be much higher than the visible answer. Cached input is far cheaper ($0.004 per million on Ember), so a stable system prompt and tool list at the start of each request pays off. Web search grounding adds $2.50 per 1,000 queries on all three.
Rate limits and data use
Rate limits are per team, not per key: 3,000 requests and 4 million tokens per minute for each Standard model, with each model counted separately. In Standard mode Meta states your content is not used for training. In Muse Code, subscriptions default to Contributor mode, so check that setting before you point Muse Code at a client repository.
Bottom line
For most apps the Meta Muse lineup comes down to three API model IDs: Wonder for the hardest work, Spark 1.3 for a cheaper main loop, and Ember for everything repetitive. Glimmer is the one to try when the model has to run on your own hardware, Wonder Full lives only in Muse Code, and Charm is a gadget. Test Ember on your real tool-calling steps first, because that is where the savings are largest.
Frequently asked questions
What is Meta Muse?
Meta Muse is Meta's brand for its current AI assistant, agent and models. For developers it covers three hosted models on the Meta Model API (Muse Wonder, Muse Spark and Muse Ember), two open-weight models (Muse Spark Open and Muse Glimmer), and products such as the Muse app, Muse Code and the Muse Charm device.
How do I use the Muse Spark API?
Create a key in the Meta Model API dashboard, then point the OpenAI SDK at https://api.meta.ai/v1 with that key and the model ID muse-spark-1.3. The same request works for muse-wonder-1.0 and muse-ember-1.0, so switching models is a model ID change.
What is the difference between Muse Ember and Muse Spark?
Muse Spark 1.3 is the mid-tier model for an agent's main loop at $1.25 input and $4.25 output per million tokens. Muse Ember is the smallest hosted model at $0.10 and $0.50, built for tool calls, routing, extraction and sub-agent tasks. Meta recommends Spark or Wonder for hard single-shot coding and knowledge-heavy questions.
Is Muse Glimmer open source and can I run it locally?
Yes. Muse Glimmer 30B is released under Apache 2.0 on Hugging Face. Its 4-bit builds target 24 GB or 32 GB of GPU memory, and Ollama offers an 18 GB build you can run with ollama run muse-glimmer:30b. It takes text and images in and does not support audio.
Can I call Muse Wonder Full through the API?
No. Meta says Muse Wonder Full is available only in Muse Code on the Power Usage and Maximum Usage plans, and no Model API model ID selects it. On the API, use muse-wonder-1.0.
Is Muse Charm a model?
No. Muse Charm is a small pocket device for talking to the Muse agent, announced at Connect 2026. It has no API or model ID, and Meta had not published a price or ship date as of 12 November 2026.
Is Meta Muse Spark 1.1 still available?
Yes. muse-spark-1.1 is still listed on the Model API at the same price as 1.3, but Meta recommends muse-spark-1.3, and the max reasoning effort only works on 1.3.
Sources
- Introducing Our Smallest, Most Efficient Model: Muse Ember (Meta AI Research, 2026-11-09)
- Pricing and rate limits (Meta Model API docs)
- Choose a model (Meta Model API docs)
- Put your ideas to work with Muse Wonder and Muse Wonder Full (Meta for Developers, 2026-10-12)
- Muse Wonder model page (Meta)
- Introducing Muse Spark 1.3 (Meta AI Research, 2026-09-02)
- SDKs and libraries (Meta Model API docs)
- Authentication (Meta Model API docs)
- Muse Glimmer 30B model card (Hugging Face)
- Muse Glimmer 30B: An On-Device Agentic Model (AI at Meta, 2026-10-19)
- muse-glimmer:30b (Ollama)
- The Biggest News From Connect 2026 (Meta Newsroom)
- Give Your Muse a Voice (Meta Newsroom, 2026-11-05)
- Day 0 support: Meta Muse Spark 1.3 (LiteLLM, 2026-09-02)
Keep reading
- Muse Spark Open Developer Guide: API, Pricing, Local SetupHow to call Meta's Muse Spark Open, what it costs on each provider, how to use it in OpenCode, and what hardware you need to run it yourself.
- LLM API Pricing Compared: Cost per Task, Not per TokenOpenAI, Claude, Gemini and DeepSeek API prices from their own pages, and a simple way to work out what one completed AI task really costs.
- GPT-6.1 API Guide: Sol, Terra, Astra and Luna ComparedGPT-6.1 Sol and Terra explained for developers: what each model is for, real API prices, cost per task, Opus 5.5 comparison and a migration checklist.
Comments
Share a question or note about this article.
Loading comments…