7/15/2026, 1:03:05 PM · foundation-models

Meta Launches Muse Spark 1.1 for Autonomous Agents with Native Sub-Agent Coordination

Meta's first model from its newly formed Superintelligence Labs arrives as a proprietary, natively multimodal reasoning system with built-in multi-agent orchestration, marking a deliberate pivot toward the agentic AI workloads where token consumption is growing fastest.

Background

Meta's Muse Spark 1.1 is the first model released by Meta Superintelligence Labs (MSL), an internal division formed in mid-2025 after a significant restructuring of Meta's artificial intelligence operations. <cite index="33-1,33-2">MSL rebuilt the company's entire AI stack from scratch — new infrastructure, new architecture, new data pipelines.</cite> The formation of MSL followed a period of competitive pressure: <cite index="33-16,33-17">by mid-2025, the prevailing narrative was that Meta had fallen behind OpenAI, Anthropic, and Google. In June 2025, Meta spent $14.3 billion to acquire a 49% non-voting stake in Scale AI and brought in co-founder Alexandr Wang as chief AI officer.</cite>

What Muse Spark Is

<cite index="28-10">Muse Spark carries "support for tool-use, visual chain of thought, and multi-agent orchestration," according to Wang, who described it as "the most powerful model that Meta has released."</cite> The model's core agentic feature is a tiered reasoning system. <cite index="33-11">It introduces three reasoning modes: Instant for casual queries, Thinking for step-by-step work, and a Contemplating mode that runs multiple sub-agents in parallel.</cite> <cite index="28-3">The Contemplating mode orchestrates multiple sub-agents to reason in parallel, positioning Meta to compete with extreme reasoning models such as Google's Gemini Deep Think and OpenAI's GPT-5.4 Pro.</cite>

In published benchmarks, <cite index="28-4">the Contemplating mode achieved 58% on "Humanity's Last Exam" and 38% on "FrontierScience Research" — figures Meta claims validate its new scaling trajectory.</cite> However, the model still trails rivals on certain agentic tasks: <cite index="33-33">on GDPval-AA, Muse Spark scores 1,427 ELO compared to GPT-5.4's 1,676, and on ARC-AGI-2 it scores 42.5 versus 76+ for top competitors.</cite>

Efficiency Architecture

<cite index="28-5,28-6">A notable aspect of the model for the company's commercial positioning is its efficiency: Meta reports that Muse Spark achieves its reasoning capabilities using over an order of magnitude less compute than Llama 4 Maverick, its previous mid-size flagship.</cite> <cite index="28-7,28-8">This efficiency is driven by a process called "thought compression," in which the model is penalized during reinforcement learning for excessive thinking time, forcing it to solve complex problems with fewer reasoning tokens without sacrificing accuracy.</cite>

Departure from Open-Source Roots

<cite index="28-20,28-21">The launch marks a controversial departure from Meta AI's "open science" roots. While the Llama series was famously accessible to developers, Muse Spark is launching as a proprietary model.</cite> This represents a meaningful strategic shift: <cite index="28-15,28-16">by early 2026 the Llama ecosystem had reached 1.2 billion downloads, averaging approximately one million downloads per day, with self-hosting offering an 88% cost reduction compared to proprietary API providers.</cite> Muse Spark does not carry open weights.

Competitive Context

The launch comes as enterprise demand for agentic workloads accelerates across the industry. <cite index="18-7,18-8">Meta had already developed Llama Stack, an interface for canonical toolchain components to build agentic applications, and identified "significant opportunities" for agentic AI systems with advanced reasoning.</cite> Competitors including <cite index="10-4">major cloud providers now offer proprietary counterparts — AWS Bedrock Agents, Microsoft Azure AI Agent Service, Google Vertex AI Agent Builder, and Anthropic's Claude with Computer Use.</cite> Third-party inference providers such as Groq and Together AI, which previously served Llama-based workloads, may be affected by Meta's move to a closed API model. <cite index="33-24,33-25,33-26">OpenAI and Anthropic are collectively valued over $1 trillion; Google's Gemini has gained traction in both consumer and developer markets; and Meta was spending $72 billion on AI infrastructure in 2025, rising to a guided $115–135 billion in 2026.</cite>

<cite index="33-21,33-22">Nine months after MSL was formed, Muse Spark is the first output. Whether it justifies the investment depends on what comes next — this model is deliberately small and fast, with larger versions already in development.</cite>

Cross-references

Sources

  1. [1]
    Agentic LLMs in 2025: How AI Is Becoming Self-Directed, Tool-Using & Autonomous | Data Science Dojo
  2. [2]
    CRAFT: Grounded Multi-Agent Coordination Under Partial Information
  3. [3]
    Agent-Testing Agent: A Meta-Agent for Automated Testing and Evaluation of Conversational AI Agents
  4. [4]
    Scaling Behavior of Single LLM-Driven Multi-Agent Systems
  5. [5]
    What is Llama and How to Use It for AI Agents | MindStudio
  6. [6]
    Agentic AI and Multiagentic: Are We Reinventing the Wheel?
  7. [7]
    Beyond Max Tokens: Stealthy Resource Amplification via Tool Calling Chains in LLM Agents
  8. [8]
    Using Strands Agents to create a multi-agent solution with Meta’s Llama 4 and Amazon Bedrock | Artificial Intelligence
  9. [9]
    GitHub - FoundationAgents/MetaGPT: 🌟 The Multi-Agent Framework: First AI Software Company, Towards Natural Language Programming
  10. [10]
    From Multi-Agent Systems and the Semantic Web to Agentic AI: A Unified Narrative of the Web of Agents
  11. [11]
    Best 10 AI Agent Frameworks for 2025
  12. [12]
    MegaAgent: A Large-Scale Autonomous LLM-based Multi- ...
  13. [13]
    MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework
  14. [14]
    Semantic Caching and Intent-Driven Context Optimization for Multi-Agent Natural Language to Code Systems
  15. [15]
    GitHub - ARUNAGIRINATHAN-K/awesome-ai-agents-2026: Awesome AI Agents for 2026 - 300+ AI Agents, Frameworks & Coding • Creative • Voice • Research • Enterprise. Comparison guides, benchmarks & deep dives · GitHub
  16. [16]
    A Review of Large Language Models and Autonomous Agents in Chemistry
  17. [17]
    Large Language Model based Multi-Agents: A Survey of Progress and Challenges
  18. [18]
    The future of AI: Built with Llama
  19. [19]
    Meta Llama 5 Open Source AI: Complete Roadmap from Behemoth to Avocado in 2026 - Sean Kim — Arts and Tech
  20. [20]
    Meta Llama: The Open-Weight AI Reaching 3.27B Users
  21. [21]
    The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation
  22. [22]
    Top 13 Agentic AI Trends to Watch in 2026
  23. [23]
    What Meta’s Latest Llama Release Means for LLM Builders in 2025
  24. [24]
    Meta Llama 2025 Ushers In the Open-Source AI Wave
  25. [25]
    deepset | deepset Accelerates Enterprise Adoption of Domain-Specific Sovereign AI with Llama Stack
  26. [26]
    Goodbye, Llama? Meta launches new proprietary AI model Muse Spark — first since Superintelligence Labs' formation | VentureBeat
  27. [27]
    Callsphere
  28. [28]
    Announcing Llama Stack Modular Agents 🦙 Streamline ...
  29. [29]
    LlamaCon 2025 and Meta’s long game in open AI | by Conrad Gray | Predict | May, 2025 | Medium
  30. [30]
    Llama Stack Released To Help Developers Build 'Agentic Apps' - The New Stack
  31. [31]
    Muse Spark vs Llama 4: Meta's Strategic Shift | WaveSpeed Blog
  32. [32]
    Agents | Developer Documentation - LlamaParse - LlamaIndex
  33. [33]
    Introducing Llama 3.1: Our most capable models to date