Meta Muse: Meta's AI Agent App and Models

Muse models (Spark, Image, Wonder, Glimmer, Spark Open)
        ↓
Meta Model API (for developers)  +  Meta AI (in Meta's apps)
        ↓
Muse app (personal agent)  ·  Muse Code (coding agent)  ·  WhatsApp, Instagram, Facebook, glasses

Standard image creation in Meta AI is free. On quality, launch coverage says Meta's internal human-rated rankings placed Muse Image second in both text-to-image and image editing, and that Meta acknowledged OpenAI's latest image model still led overall. Those are Meta's own evaluations, not independent ones.

Muse Wonder: built for agents

Muse Wonder is Meta's most capable model so far, and it's aimed squarely at agents. Meta highlights two abilities. Active Memory keeps important context about you across conversations and pulls the relevant details into the current task. Active Focus keeps the model working toward hard goals: it tries other approaches when blocked, asks for clarification when it needs to, and checks in before important actions.

Meta claims state-of-the-art results on agent, memory and coding benchmarks, but independent results were thin at the time of writing. Muse Wonder is in preview on the Meta Model API as muse-wonder-1.0. The larger Muse Wonder Full is limited to Muse Code's higher subscription plans.

Muse Glimmer and Muse Spark Open: the ones you can run yourself

These two are very different propositions, and the table makes the choice fairly obvious:

Detail	Muse Glimmer	Muse Spark Open
Size	30B parameters	816B total, 42B active (mixture of experts)
License	Apache 2.0	Muse Model License (limited commercial terms)
Hardware	One consumer GPU or Mac, about 24 to 32 GB memory at 4-bit	Data-center GPUs, or two 512 GB Mac Studios using MLX
Best for	Local agents, function calling, local coding	Coding and agentic work at frontier scale
Context	Not stated by Meta	1M tokens

Both are on Hugging Face, and Muse Spark Open is also available through major cloud model catalogs and hosted inference providers. If you want something running on your own machine this week, Glimmer is the realistic option, and its Apache 2.0 license is the only genuinely permissive one in the Muse family. If you haven't run a model locally before, Local AI Inference in .NET covers running models on Windows without cloud APIs.

Calling Muse models from code

The Meta Model API uses the base URL https://api.meta.ai/v1 with a bearer API key. Meta's own examples point the OpenAI SDK at that URL, and the API supports Chat Completions, Responses and Messages-style requests for Muse Spark and Muse Wonder, plus image generation and edit endpoints for Muse Image. The model IDs you'll use:

• muse-spark-1.3, recommended for new work, in Standard mode
• muse-spark-1.3-contributor, cheaper, but your content may be used to train Meta's models
• muse-wonder-1.0, preview, Standard mode only
• muse-image-1.0, for image generation and edits
• muse-voice-transcribe-1.0, for speech to text

The Standard/Contributor split matters more than the model choice. Standard mode means your content is not used to train Meta's models; Contributor mode trades that away for a lower price. For anything that touches customer or company data, use Standard.

In Python, it's the OpenAI client with a different base URL:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key=os.environ["MODEL_API_KEY"],
)

response = client.chat.completions.create(
    model="muse-spark-1.3",
    messages=[
        {"role": "system", "content": "You are a concise technical assistant."},
        {"role": "user", "content": "Explain dependency injection in ASP.NET Core in 3 bullets."},
    ],
)

print(response.choices[0].message.content)

The same works from C# with the OpenAI .NET SDK:

using System.ClientModel;
using OpenAI;
using OpenAI.Chat;

var client = new ChatClient(
    model: "muse-spark-1.3",
    credential: new ApiKeyCredential(Environment.GetEnvironmentVariable("MODEL_API_KEY")!),
    options: new OpenAIClientOptions { Endpoint = new Uri("https://api.meta.ai/v1") });

ChatCompletion completion = await client.CompleteChatAsync(
    new SystemChatMessage("You are a concise technical assistant."),
    new UserChatMessage("Explain dependency injection in ASP.NET Core in 3 bullets."));

Console.WriteLine(completion.Content[0].Text);

Third-party pricing trackers list Muse Spark 1.3 at about $1.25 per million input tokens and $4.25 per million output tokens, and new accounts were reported to get free credits. Check Meta's pricing page before you budget.

The limits and the open questions

Availability is the first constraint. The agent app is US-only and adults-only, model availability on the API varies by country, and Muse Wonder isn't offered in South Korea. Meta AI powered by Muse Spark is available much more widely than the agent.

Data use is the second. Consumer queries are reportedly used for training unless you opt out, and on the API, Contributor mode allows training too. Meta isolates each agent in its own VM and gates access with permissions, but the sensible approach is still to grant only what a task needs and review anything that involves a purchase.

Then there's the reception from the rest of the web. Reports say Amazon blocked Muse from its retail site on September 21, 2026, on the grounds that the agent didn't identify itself; Meta says Muse has no visibility into passwords or payment methods. Expect other sites to block AI agents as well. 404 Media also reported that some beta phone calls were handled by human call-center staff, which Meta called standard internal testing.

Licensing is easy to get wrong. Muse Spark and Muse Wonder are closed models, and Muse Image is closed too. Muse Spark Open has downloadable weights but uses Meta's own license with limited commercial terms. Only Muse Glimmer is Apache 2.0.

Finally, the headline benchmarks are mostly Meta's own. Treat them as claims until independent tests confirm them.

How Muse compares with other AI tools

Meta's biggest advantage is distribution. Muse is built into WhatsApp, Instagram, Facebook and Meta's glasses, which billions of people already use, while ChatGPT, Gemini and Claude mostly live as standalone assistants. Its biggest open question is trust, because an agent like Muse needs deep access to your accounts before it's useful.

On the specifics, I'd compare along the lines that actually differ. For agent work, look at each product's permission model and which sites it can operate on, since a dedicated VM per user is only as useful as the sites that let it in. For images, Muse Image is free and has good annotation-based editing, but by Meta's own account OpenAI's and Google's image models still lead in some areas. For the API, compare price per token, data-retention terms and regional availability rather than context length alone. For open weights, Glimmer's Apache 2.0 license is the standout, and Spark Open needs a careful read of its license and serious hardware. If you're weighing open models more broadly, our roundup of open-source LLMs worth experimenting with is a good next stop.

For developers, my starting point would be muse-spark-1.3 in Standard mode through the OpenAI SDK you probably already use. It costs one changed base URL to find out whether it fits your workload.