GGUF Discovery

Blog & Guides

Back to All Articles

Meta Muse Wonder & Wonder Full: Technical Deep Dive

October 14, 2026 By Hussain Nazary Technical Analysis

Meta Muse Wonder & Wonder Full: A Complete Technical Deep Dive

On October 12, 2026, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full—Meta's first true frontier AI models. Trained at the gigawatt-scale Prometheus data center, these models represent Meta's return to the top tier of AI development, just 14 months after a comprehensive reorganization. This deep dive examines everything Meta has publicly shared about the architecture, training, capabilities, and positioning of the Muse Wonder family.

What Is Meta Muse Wonder?

Meta Muse Wonder is a frontier-class large language model family designed specifically for personal and professional agentic tasks. Released on October 12, 2026, the family includes two distinct variants:

Both models support a 1,048,576-token context window and accept multimodal inputs (text, images, audio, video, and PDF files) while outputting text only. They are the first models from Meta Superintelligence Labs (MSL), the division formed after Meta's August 2025 AI reorganization under Chief AI Officer Alexandr Wang.

"Trained at our gigawatt-scale Prometheus data center, Muse Wonder and Muse Wonder Full bring together the work of teams across Meta to deliver frontier capabilities for complex tasks at work and in everyday life—another step toward bringing personal superintelligence to the billions of people who use our products."
— Meta AI Research Blog

The Two Variants: Access and Positioning

Feature Muse Wonder Muse Wonder Full
Model ID muse-wonder-1.0 muse-wonder-full-1.0
API Access Meta Model API Not available via API
Context Window 1,048,576 tokens 1,048,576 tokens
Input Pricing $4.00 per 1M tokens Subscription-only
Output Pricing $20.00 per 1M tokens Subscription-only
AI Intelligence Index 52 56
Coding Agents Index 63 68

The most important distinction is not simply capability—it's how developers can access each model. Muse Wonder is available through a standard metered API suitable for custom applications. Muse Wonder Full is only accessible through eligible Muse Code subscriptions, with no public API endpoint.

Architecture and Scale

Parameter Count: ~3 Trillion Total

While Meta has not officially disclosed the exact parameter count, industry sources estimate Muse Wonder has approximately 3 trillion total parameters. According to industry analysis, channel checks indicate Wonder is smaller than both GPT-6 Astra and Claude Fable 5.1, but still firmly in the frontier model category.

"Our channel checks indicate that Wonder has roughly 3T total parameters, making it smaller than both Astra and Fable."
— Industry Analysis Report, October 2026

This places Muse Wonder in a competitive position:

Prometheus Data Center: Gigawatt-Scale Infrastructure

Both Muse Wonder variants were trained at Meta's Prometheus data center, described as a "gigawatt-scale" facility located in central Ohio. This represents one of the largest AI training installations in the world.

Key infrastructure details:

Training Methodology: From Spark to Wonder

Meta's approach to developing Wonder followed a deliberate, multi-generation strategy:

  1. Recipe Validation: The team validated training recipes, data mixtures, and RL environments on small-scale models
  2. Incremental Scaling: Each generation increased model size by 4-5x
  3. Spark Foundation: Muse Spark (codenamed "Avocado") served as the foundation, with pretraining beginning in late 2025
  4. Wonder Training (codenamed "Watermelon"): Planning began in August 2025, culminating in the October 2026 release

This represents faster iteration than Meta achieved during the Llama era. The transition from planning to frontier-competitive model took just over 14 months—a remarkable pace in frontier AI development.

Multimodal Architecture

Muse Wonder is designed as a natively multimodal system, accepting:

However, the model outputs text only. For image and video generation, Meta offers separate products (Muse Image and Muse Video).

What Meta Has Not Disclosed: Architecture Specifications

Meta has kept the detailed architecture specifications of Muse Wonder proprietary. Unlike open-weight models where technical details are published, the following information remains undisclosed:

This is a deliberate strategic choice by Meta. While the company open-sources its Llama model family with full technical specifications, Muse Wonder and Wonder Full are positioned as proprietary commercial models competing directly with OpenAI's GPT series and Anthropic's Claude family.

Why this matters: Without architecture details, researchers cannot reproduce the model, fine-tune it for specialized tasks, or run it on custom infrastructure. The only way to use Muse Wonder is through Meta's controlled API or Muse Code subscription, giving Meta complete control over pricing, access policies, and deployment.

The limited public information consists of:

This contrasts sharply with models like Qwen, Mistral, and Llama, where detailed architecture specifications, model weights, and technical papers are publicly available. Even closed models from competitors like Anthropic occasionally publish architectural insights in research papers.

Meta's approach with Muse Wonder mirrors OpenAI's GPT-4 strategy: share high-level capabilities and benchmark results, but keep the architectural recipe proprietary to maintain competitive advantage.

Core Capabilities: Active Memory and Active Focus

Meta describes two flagship capabilities that define the Muse Wonder family: Active Memory and Active Focus. These represent specific architectural and training decisions that differentiate Wonder from conventional chatbots.

Active Memory: Context That Persists

Active Memory is Meta's term for the model's ability to:

Traditional conversational AI loses important details as tasks grow longer. Active Memory addresses this by training the model to maintain and surface relevant context at the right moments.

Example scenario: An AI travel agent needs to combine a user's dietary preferences (mentioned two weeks ago) with a new calendar event, transportation constraints, and updated destination information—all while respecting a previously stated budget.

On Meta's internal MemoryBench evaluation:

Active Focus: Persistent Execution Across Complex Tasks

Active Focus describes Wonder's ability to:

This capability is especially critical for long-running AI agents. Consider an autonomous coding task:

  1. Inspect repository structure
  2. Identify dependencies and constraints
  3. Modify multiple files
  4. Run test suite
  5. Diagnose test failures
  6. Revise implementation
  7. Re-test until passing

A model that produces convincing initial code but can't respond to failed tests has limited practical value. Active Focus trains Wonder to persist through blockers instead of stopping and handing the task back to the user.

Benchmark Performance: How Capable Is Muse Wonder?

Artificial Analysis Intelligence Index

The most comprehensive independent evaluation comes from Artificial Analysis, which released its Intelligence Index v4.3 shortly after Wonder's launch:

Model Intelligence Index v4.3
Claude Opus 5.5 58
Muse Wonder Full 56
Claude Sonnet 5.5 56
GPT-6 Astra 53
Gemini 4 Argon 53
Muse Wonder 52
Muse Spark 1.3 48

Key takeaway: Muse Wonder Full ranks second globally, tied with Claude Sonnet 5.5 and just 2 points behind Opus 5.5. Importantly, Wonder Full improved over Spark 1.3 across all 10 evaluations in the Intelligence Index.

Coding Agents Index: Elite Performance

In agent-driven coding workflows, Muse Wonder Full reaches the top tier:

Model Configuration Coding Agents Index
Muse Wonder Full in Muse Code 68
Claude Sonnet 5.5 in Claude Code 68
Claude Opus 5.5 in Claude Code 66
Muse Wonder in Muse Code 63
Muse Spark 1.3 in Muse Code 54

Wonder Full matches Claude Sonnet 5.5 and exceeds Opus 5.5 in the evaluated configurations.

Long-Horizon Engineering Tasks

Muse Wonder Full excels particularly in sustained software engineering tasks:

Benchmark Wonder Full Opus 5.5 Astra
SWE-Marathon 57.5 46.3 42.5
DeepSWE 79.0 ~74.0 ~74.0
Terminal-Bench 4.0-AA 57.1 ~59.5 ~59.5

Wonder Full shows a substantial lead on marathon-length coding tasks, reinforcing its positioning as an agent-first model.

Professional Domain Expertise

Meta designed Wonder for knowledge-intensive professional work in law, finance, and healthcare:

Benchmark Wonder Full Opus 5.5 Astra
Harvey Legal Agent Benchmark 13.1 4.2 8.6
JobBench 75.3 ~75.0 —
HealthBench Pro 62.3 ~65.0 ~65.0

Wonder Full shows exceptional performance on legal agent work, more than doubling Astra and tripling Opus 5.5.

Notable Weaknesses: General Knowledge Accuracy

Muse Wonder's strongest performance does not extend uniformly across all evaluations. On AA-Omniscience, a general knowledge benchmark, Wonder Full scores 30—below several competing frontier models.

Additionally, Artificial Analysis reports a benchmark-specific hallucination rate of approximately:

While this represents improvement, for research and factual questions, retrieval augmentation and citation verification remain essential.

Pricing and Access Economics

Muse Wonder API Pricing

Meta publishes transparent usage-based pricing for standard Muse Wonder through the Meta Model API:

Usage Category Price
Input tokens $4.00 per 1M
Cached input tokens $0.20 per 1M
Output tokens $20.00 per 1M
Web search grounding +$2.50 per 1,000 searches

This places Muse Wonder in the premium frontier-model category, roughly equivalent to Claude Opus 5.5 at $4/$20 per million tokens.

Real-World Cost Examples

Example 1: Standard Analysis Request

Request processes 100,000 uncached input tokens and generates 10,000 output tokens:

Example 2: Multi-Step Agent Workflow

Agent uses 300,000 uncached input, 800,000 cached input, and 80,000 output tokens:

Example 3: Monthly SaaS Application

100 requests per day, averaging 30,000 uncached input and 5,000 output tokens per request:

Muse Wonder Full: Subscription-Only Access

Muse Wonder Full has no published per-token API pricing. It's distributed exclusively through Muse Code subscriptions:

Plan Monthly Price Usage Allowance Wonder Full Access
Everyday Usage Free Baseline ❌ No
Power Usage $50 20x Everyday ✅ Yes
Maximum Usage $200 80x Everyday ✅ Yes

Meta documents five-hour and weekly usage resets for these subscriptions.

Meta Superintelligence Labs: The Team Behind Muse Wonder

The August 2025 Reorganization

In June 2025, Meta paid $14.3 billion for a 49% stake in Scale AI and brought in Alexandr Wang as Chief AI Officer. This marked the beginning of a comprehensive reorganization:

The hiring drew skepticism at the time: Meta could pay to attract talent, but whether it could build a frontier lab remained an open question.

From Skepticism to Success

Just over a year later, MSL delivered Muse Wonder—a model that scores 56 on the independent Intelligence Index, tied with Claude Sonnet 5.5 and second only to Opus 5.5. According to industry analysts:

"In our view, this represents faster iteration than Meta achieved in the Llama era. We believe Meta is back among the leading frontier AI labs, and we expect it to stay there."
— Industry Analysis, October 2026

The team first validated its training recipe, data, and RL environments on small models, then increased model size by 4-5x with each generation. Muse Spark began pretraining at the end of 2025. Muse Wonder planning began in August 2025, culminating in the October 2026 release.

How to Use Muse Wonder: Developer Guide

Step 1: Obtain API Access

  1. Visit the Meta Developer Platform
  2. Create an API key through the supported account workflow
  3. Store the key securely as an environment variable (MODEL_API_KEY)

Step 2: Install the OpenAI-Compatible SDK

The Meta Model API supports OpenAI-compatible SDK integrations. For Node.js/TypeScript:

npm install openai

Step 3: Make Your First Request

import OpenAI from 'openai';

const client = new OpenAI({
  baseURL: 'https://api.meta.ai/v1',
  apiKey: process.env.MODEL_API_KEY,
});

async function callMuseWonder() {
  const response = await client.chat.completions.create({
    model: 'muse-wonder-1.0',
    messages: [
      {
        role: 'user',
        content: 'Explain Active Memory in Muse Wonder.'
      }
    ],
    max_tokens: 1000,
  });

  console.log(response.choices[0].message.content);
}

callMuseWonder();

Step 4: Advanced Capabilities

Meta's API documentation describes support for several agent-focused capabilities:

Rate Limits

The API supports:

Best Practices for Building Agents with Muse Wonder

1. Define Verifiable Acceptance Criteria

Instead of vague goals like "improve the application," specify:

2. Separate Planning and Execution

For consequential operations (deployments, destructive changes, financial transactions), require an implementation plan before execution. Human approval checkpoints reduce the risk of completing the wrong task.

3. Use Prompt Caching for Repeated Context

Large repositories and multi-agent workflows often reuse substantial context. Caching repeated prompt prefixes can reduce costs by 90% on cached portions.

4. Route by Task Complexity

Not every request justifies a premium frontier model:

5. Use Persistent Checkpoints

Long-running agents should store task progress outside the prompt:

Privacy and Data Usage

Meta Model API: Standard Mode

According to Meta's published policies, Standard mode does not use submitted prompts, code, files, context, or model responses to train future Meta models. Regular Muse Wonder is available through the public API in Standard mode only.

Muse Code: Contributor Mode

Muse Code subscriptions default to Contributor mode, which allows Meta to use submitted content (including code and model responses) for future model training. Subscribers can turn Contributor mode off in settings.

Important for enterprise users: The distinction matters for proprietary repositories, customer data, internal documents, and confidential business processes.

The Competitive Landscape

The Four-Way Frontier Contest

According to industry analysis, a year ago the frontier AI group consisted mainly of OpenAI and Anthropic. In 2026, Google rejoined this tier, followed by Meta, and a four-way contest is taking shape:

Lab Flagship Model AI Index v4.3 Key Strength
Anthropic Claude Opus 5.5 58 Overall capability leader
Meta Muse Wonder Full 56 Agent tasks, legal work
Anthropic Claude Sonnet 5.5 56 Coding agents
OpenAI GPT-6 Astra 53 Cost efficiency
Google Gemini 4 Argon 53 Multimodal reasoning

Compute Bottleneck: The Real Constraint

According to media reports, in July 2026 Meta was preparing to rent out surplus GPUs and was in talks with Anthropic. By October, its own needs had forced it to shelve the plan. The compute shortage is expected to ease only once new infrastructure ships in 2027.

Future Outlook

Industry analysts expect Meta to remain in the top tier of frontier AI labs. The team has proven itself, the Prometheus expansion and Hyperion build-out will add compute, and Meta's existing apps provide established distribution channels.

Progress on the next model depends mainly on when new capacity comes online. With the infrastructure investments underway and the proven MSL team, Meta appears positioned to continue competing at the frontier.

Conclusion

Meta Muse Wonder and Wonder Full represent a significant milestone: Meta's return to frontier AI development after a comprehensive reorganization. With approximately 3 trillion parameters, trained at the gigawatt-scale Prometheus data center, these models demonstrate that Meta Superintelligence Labs can compete with OpenAI, Anthropic, and Google.

The flagship capabilities—Active Memory and Active Focus—position Wonder as an agent-first model optimized for long-running tasks rather than single-turn responses. Benchmark performance backs this up, with Wonder Full scoring 56 on the Intelligence Index (second globally) and 68 on the Coding Agents Index (tied for first).

For developers, the choice between Wonder and Wonder Full comes down to access requirements: Wonder offers API-based integration at premium pricing ($4/$20 per million tokens), while Wonder Full requires a Muse Code subscription but delivers higher capability.

Just 14 months after skeptics questioned whether Meta could build a frontier lab, MSL has delivered a model that ranks among the world's best. The speed of iteration, quality of execution, and competitive positioning all suggest Meta will remain a major player in the ongoing frontier AI race.

Sources: Meta AI Research, AI Aide List, Artificial Analysis, Industry Reports

Connect on Instagram