Meta Muse Spark 1.3: a technical deep dive into architecture, benchmarks, API and how it compares

01From Llama to Muse: why Meta changed course

For years, Meta was synonymous with open models. From 2023 to 2025 it released four generations of Llama. (Wikipedia) In 2026 it changed course: the Muse Spark line from the new Meta Superintelligence Labs division replaced Llama, and it is closed.

Muse Spark timeline

The first Muse Spark in April scored 52 on the Artificial Analysis Intelligence Index and trailed the frontier (57 for GPT-5.4 and Gemini 3.1 Pro, according to Flowtivity). It stood out for efficiency, though: it ran the whole index with 58 million output tokens, while Claude Opus 4.6 used 157 million. (Artificial Analysis, Flowtivity)

What about open weights?

Muse Spark 1.3 is not open. In its release notes, Meta lists open weights for Muse Spark only as a future roadmap item. If you need a downloadable Meta model today, there is Muse Glimmer: a dense 30 billion parameter multimodal model distilled from Muse Spark, under the Apache 2.0 license, with a default context of 128,000 tokens. (TrueFoundry)

02What Muse Spark 1.3 brings over 1.2

Version 1.3 is mainly an efficiency and behavior upgrade, not a leap in general intelligence. According to Meta's engineers, it makes about 20% fewer tool calls on the same work and uses about 25% fewer tokens. (TrueFoundry)

Why efficiency matters

25% fewer tokens on the same work means a 25% lower bill before any other optimization. It is Meta's internal measurement, though, not independent verification; measure the real savings on your own traffic. (TrueFoundry)

Watch out for audio

Meta's own documentation says audio understanding is not fully supported in 1.3 and response quality may be worse. For audio it recommends staying on 1.2 or using the separate Muse Voice Transcribe model. (TrueFoundry)

03Specs and architecture

Muse Spark 1.3 is a natively multimodal model with a context of over a million tokens, available only through the API.

What we know about the architecture

Meta has not published the parameter count or architecture details. We know the model is trained on agentic workflows and competitive programming and has native multimodal perception. (note.com (kazu), ExplainX)

Limits are shared across a team

All API keys on a team share one limit of 3,000 requests and 4 million tokens per minute. A separate key per service does not split the capacity, so one batch job can block an interactive product. Meta recommends watching the x-ratelimit-remaining-tokens header and slowing down before you hit the limit. (TrueFoundry)

04Muse Spark 1.3 benchmarks in depth

Muse Spark 1.3 leads on coding and long context, but on agentic tasks it loses to Claude Opus 5 almost everywhere. Meta compares its max setting with Muse Spark 1.2 (xhigh), GPT-5.6 Sol (max) and Claude Opus 5 (max). The numbers come from Meta's table (Meta: Muse Spark).

Coding

Long context

Agentic tasks and knowledge work

Max vs xhigh: a big gap on agentic tasks

The results above are for the max level, which was in preview at launch and is now available in Muse Code and the Meta Model API (Standard tier only). The lower xhigh level is noticeably weaker on agentic tasks:

Three catches in Meta's comparison

1. Different effort levels. Version 1.3 and the competitors ran at max, while 1.2 ran only at xhigh. Part of the jump between versions is a settings difference, not a model difference.
2. Different OSWorld versions. All models ran on OSWorld 2.0 version 08.08, except Muse Spark 1.2 on version 06.24.
3. Best available number. For each model, Meta takes the highest comparable figure from its own measurement, a public leaderboard or the vendor. It is not one controlled run.

Source: TrueFoundry, based on Meta's methodology. Meta itself admits that its tests of competing models may not reflect their best performance. (Meta: methodology)

05How it compares with the September 2026 flagships

On DeepSWE, Muse Spark 1.3 is the best of all September models by the vendors' numbers; on the independent Artificial Analysis index it sits mid-pack, and on price it is among the cheapest.

Benchmarks vs reality

The 75.4% DeepSWE result is not yet on the public DeepSWE leaderboard; only the older Muse Spark 1.2 is there, at 55%. On the Artificial Analysis index v4.3, Muse Spark 1.3 (max) scores 48; ahead of it are Claude Opus 5.5 (58), Claude Fable 5.1 and GPT-6 Astra (both 53), Claude Opus 5 (51) and Claude Fable 5 (50). (Artificial Analysis, as of September 26, 2026) In a hands-on game-building test, it fell behind models with lower scores. (MindStudio)

For more on the competitors, see our GPT-6 Astra guide, Claude Fable 5.1 guide, Gemini 3.8 Flash guide and Grok 4.7 guide.

For all seven September models side by side, with pricing and benchmark charts, see our big comparison of September 2026 AI models.

06Where to find Muse Spark 1.3

Muse Spark 1.3 is available in four channels, all aimed at developers.

In the available sources, we found no confirmation that Muse Spark 1.3 powers the consumer Meta AI assistant in WhatsApp, Instagram or Messenger.

Muse Code: a default with a catch

According to DataCamp, Muse Code switches you to muse-spark-1.3-contributor at startup, so Meta trains on your code until you change it. (DataCamp) The official Muse Code documentation still lists muse-spark-1.2 as the default model. For company or client code, check that you are on the Standard tier right after installation.

07For developers: API and integration

Moving from 1.2 takes a single change: the new model ID. Endpoints, SDKs and prices stay the same. (DataCamp)

Compatibility

The Meta Model API supports three formats: Responses, Chat Completions and Messages. Most existing clients for other models only need a new base URL and model ID, with no code rewrite. (TrueFoundry)

An illustrative call

An example using a Chat Completions compatible client. Fill in the base URL and the key variable name from Meta's developer documentation:

import os
from openai import OpenAI

client = OpenAI(
    base_url="<base URL from the Meta Model API docs>",
    api_key=os.environ["META_API_KEY"],
)

response = client.chat.completions.create(
    model="muse-spark-1.3",  # Standard: Meta does not train on your data
    messages=[
        {"role": "user", "content": "Go through the repository and find the cause of the failing test."}
    ],
)

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

Deployment checklist

1. Decide on the tier up front, in writing. Standard for anything that touches customer data; Contributor only where training on the data is genuinely acceptable.
2. Use a stable internal model name. Meta ships new versions every four weeks; moving to 1.4 should be a config change, not a deployment to every service.
3. Set the reasoning level on purpose. Max is only on Standard and brings both the best results and the highest bills. Start lower and raise it by task type.
4. Budget for search. Web search grounding costs $2.50 per 1,000 queries on top of tokens.
5. Watch the shared limit. Read the x-ratelimit-remaining-tokens header and slow down in time.
6. Send only part of your traffic first. Compare the cost per acceptable answer with your current model and expand only where it wins.

Sources: TrueFoundry, Meta: pricing and limits

08Pricing and license

Meta offers two tiers: Standard at $1.25 / $4.25, and Contributor, more than 90% cheaper, where Meta trains on your data.

What makes the pricing stand out

• No long-context surcharge. The same price per token whether the window is nearly empty or full. Grok 4.7 doubles its price from 200,000 tokens in the prompt (Grok 4.7 guide); Muse Spark does not.
• Contributor is 92% cheaper on input and 95% cheaper on output. Meta openly says the discount is in exchange for consent to use your prompts and responses for training.
• Search is billed separately: $2.50 per 1,000 queries.

The “about $0.10 per million” figure that appeared in some roundups is not a flat rate but a blended Contributor price. With an 8 : 1 input to output ratio, the price list works out to about $0.11 per million tokens.

License

The model is proprietary and available only through the API. (Emergent) Meta's only open model right now is the smaller Muse Glimmer under the Apache 2.0 license (text and images).

09Privacy and safety

The biggest risk with Muse Spark 1.3 is not the model but the tier: one suffix in the model ID decides whether customer data becomes Meta's training data.

The Contributor tier risk

A discount of over 90% is exactly the kind of saving a developer reaches for when cutting costs without reading the terms. The two tier IDs differ only by the -contributor suffix. If a production service moves to Contributor by mistake, customer prompts become training data and nothing visibly changes in the app. Enforce the rule centrally, not in each repository. (TrueFoundry)

Where the data lives

Closed weights mean prompts always go to Meta's API. The option to run the model on your own infrastructure, which the open Muse Glimmer offers, does not exist here. For regulated data, consult your lawyers. (TrueFoundry)

Model safety

• Higher resistance to harmful inputs and prompt injection.
• Better calibration on irreversible actions; it asks for confirmation before major steps.
• It admits what it cannot do instead of making up a result.
• At launch, the max level was announced only after further safety testing was complete.

Sources: TrueFoundry. We did not find a separate system card with safety test results in the available sources.

10Weaknesses, criticism and open questions

Muse Spark 1.3 has strong numbers, but also a number of caveats worth knowing before you deploy it.

1. Unverified headline score. The 75.4% DeepSWE result is not yet on the public DeepSWE leaderboard; it is Meta's measurement. (Tenten)
2. A gap between numbers and practice. In a hands-on game-building test, it fell behind models with lower scores. (MindStudio)
3. Best results only at max. The xhigh setting is noticeably weaker on agentic tasks (OSWorld 57.2% vs 66.9%). (MarkTechPost)
4. Behind the competition on agentic tasks. Claude Opus 5 leads on GDPval-AA, JobBench, OSWorld (partial) and AutomationBench, and GPT-5.6 Sol leads on DeepSearchQA and instruction following. Muse Spark 1.3 leads only on the strict OSWorld variant. (Meta: Muse Spark)
5. Degraded audio. Audio understanding is not fully supported in 1.3.
6. Fast version churn. Three versions between April and September, with only four weeks between 1.2 and 1.3. You will change a hard-coded model ID more often than you expect. (TrueFoundry)
7. Closed. No open weights, no self-hosting and no architecture details.
8. Training on your code by default in Muse Code. If you do not change the setting, you are providing your code for training.

Open questions

• Will the DeepSWE result appear on the public leaderboard, and will it hold up?
• When, and in what form, will Meta open the Muse Spark weights, as its roadmap promises?

11Use cases: when to deploy Muse Spark 1.3 and when not to

Muse Spark 1.3 makes the most sense for long coding tasks and giant-context work at a low price.

When to use it

• Analyzing a large repository. 59.4% on SWE-Atlas Codebase QnA and a million-token context with no surcharge; the whole codebase fits in one request.
• Long agentic coding. The highest reported DeepSWE score and a quarter fewer tokens than 1.2.
• Search across huge documents. 98.1% on MRCR in the 512K to 1M token range.
• Routing between several models. A cheap fourth option next to OpenAI, Anthropic and Google for long agentic runs.
• Internal prototypes and synthetic data. The Contributor tier at a fraction of the price, where training on the data does not matter.

When not to

• Audio processing: use Muse Spark 1.2 or Muse Voice Transcribe.
• General agentic tasks and computer use: Claude Opus 5 and GPT-6 Astra lead.
• You need open weights or self-hosting: choose Muse Glimmer or another open model.
• Sensitive data on the Contributor tier: never.
• Short answers and simple edits: a 75.4% score is no reason to migrate services that only handle short tasks. (Tenten)

12Prompting tips and best practices

Muse Spark 1.3 is trained to collaborate in long threads, so you get the most out of it with a clear goal and ongoing steering.

1. One thread, several tasks. Version 1.3 is better at telling which task a new instruction belongs to. Still, name your tasks (“Task A: refactor, Task B: tests”) so they do not get mixed up.
2. Let it ask questions. The model asks about ambiguities and asks for confirmation before major steps. Do not just answer “just do it”; a short reply saves expensive mistakes.
3. Ask for brevity. Version 1.3 writes less verbose code. Reinforce that with an instruction such as “no unnecessary comments or helper files.”
4. Use the full context. Load the whole repository or documentation at once; you do not pay a long-context surcharge.
5. Match the reasoning level to the task. Use xhigh or lower for routine work, and max only for the hardest agentic tasks.
6. Ask about limits. For uncertain tasks, add “if you cannot do this, tell me why”; the model is trained to admit what it cannot do.

These practices follow from the description of the changes in 1.3 (TrueFoundry, note.com (kazu)).

13Muse Spark 1.3 FAQ

14Verdict: is Muse Spark 1.3 worth it?

Muse Spark 1.3 is a quiet but useful release: fewer tokens, fewer mistakes in long threads and a million-token context at a low price. For teams that code with agents and work with large repositories, it is a serious candidate next to models from OpenAI, Anthropic and Google.

It comes with caveats, though. The best numbers come from Meta and from the max level, audio got worse, the weights are closed and you pay for the cheapest tier with your data. Test it on your own tasks on the Standard tier, set the Contributor rules centrally and then decide.

Want to put AI agents to work in your own processes? We help you choose the model and connect it to your systems as part of our AI implementation service.

15Sources

• Meta: Muse Spark (model and benchmarks)
• Meta: API models
• Meta: pricing and limits
• Meta AI Research: introducing Muse Spark 1.3
• Meta: Muse Spark 1.3 evaluation methodology
• Artificial Analysis: Intelligence Index v4.3
• TrueFoundry: Muse Spark 1.3, what Meta shipped
• DataCamp: a hands-on Muse Spark 1.3 guide
• DataCamp: features, benchmarks and pricing
• MarkTechPost: Meta releases Muse Spark 1.3
• AIToolsReview: Muse Spark 1.3 review
• ExplainX: Muse Spark 1.3 vs Opus 5
• Tenten: the 75.4% DeepSWE score
• MindStudio: why the numbers do not add up
• Flowtivity: benchmarks and agents
• note.com (kazu): Muse Spark 1.3 explained
• Emergent: Meta launches Muse Spark 1.3
• Wikipedia: Meta Superintelligence Labs