Meta's Muse Spark takes the stage to revolutionize AI

In short : Muse Spark is the very first artificial intelligence model developed by Meta Superintelligence Labs. Recently unveiled by Meta, this model marks the launch of its new family named Muse. Designed to compete head-to-head with leading proprietary models such as OpenAI's GPT and Anthropic's Claude Opus, Muse Spark is firmly part of today's technology race.

‍

Muse Spark: what is it?

Muse Spark is the first large multimodal reasoning model developed by Meta Superintelligence Labs, an internal Meta division created to fundamentally rethink the company's artificial intelligence (AI) architecture and strategy. Unlike the Llama series, which relies on a parameter “stacking” approach and is largely open to the community, Muse Spark represents a major break. This model is closed and designed primarily as an internal engine for Meta's products and services, before any prospect of opening up to the public.

‍

Meta's origins and vision

Muse Spark is the result of a strategic reorganization of Meta in the AI field. In 2025, faced with mounting delays and performance deemed insufficient for the Llama models compared with those of OpenAI, Anthropic and Google, Mark Zuckerberg decided to launch Meta Superintelligence Labs. This new entity was tasked with completely rebuilding the company's training and modeling stack.

‍

Under the leadership of Alexandr Wang, former founder of Scale AI, an internal team nicknamed “Avocado” worked for nearly nine months to develop a model designed from the ground up for assisted superintelligence. The goal was to create an assistant capable of solving complex problems in fields such as science, mathematics and health. Muse Spark thus marks a strategic shift, moving from an openness-centered approach to a priority placed on performance and control of the model.

‍

Key features of Muse Spark

Muse Spark is a natively multimodal model, able to process text, images and audio simultaneously within a single architecture, without needing to add external modules. It also natively integrates the concept of “tool use,” that is, the ability to call external tools (such as APIs, computation engines or databases) to answer complex queries.

‍

Another key feature is the “visual chain-of-thought,” which allows the model to visually break down a problem posed through an image, showing its reasoning step by step. This strengthens transparency and helps users better understand the answers provided.

‍

Finally, Muse Spark implements an orchestration of reasoning agents that work in parallel to synthesize their results. This approach makes it possible to handle complex tasks without endlessly extending a single agent's thinking time, which improves the overall efficiency of the system.

‍

An open but proprietary model: distribution strategy

While Meta became known for its largely open-source Llama models, Muse Spark takes the opposite approach. It is a proprietary model whose weights cannot be downloaded and which cannot be self-hosted. Currently, Muse Spark is accessible through Meta AI on the web and the Meta AI app, as well as gradually across other group services such as Facebook, Instagram, WhatsApp and Ray-Ban AI.

‍

For external developers, a private preview of the API is offered, but by invitation only, which considerably limits access to this technology. This strategic choice aims to keep full control over the central engine of Meta's AI ecosystem, while reserving the commercial and experimental benefits for the company itself.

‍

In addition, Meta has announced its intention to gradually reopen certain versions or branches of Muse Spark. However, no precise timeline or specific model has been communicated, which sustains a degree of ambiguity about the real scope of this potential long-term opening.

‍

Why Muse Spark promises to revolutionize artificial intelligence

Muse Spark does not merely match the market's leading models: it redefines the rules of the game by combining power, versatility and efficiency. Thanks to an architecture designed for multimodal reasoning and agent-based operation, it paves the way for uses that are more natural, faster and economically viable, whether for individuals or businesses.

‍

Model performance and speed

Muse Spark stands out for its ability to produce very high-quality answers while maintaining competitive response times. Early analyses show that it uses significantly fewer tokens than many models at its level, which translates into lower compute costs and reduced latency for the end user.

‍

Its ability to “think more with less” allows it to remain high-performing even in high-load contexts, such as real-time chatbots or assistants embedded in mass-market applications.

‍

The various operating modes, Instant, Thinking and Contemplating, give the user precise control over the trade-off between speed and depth of reasoning. Unlike other models that simply impose a longer thinking time, Muse Spark can, in its Contemplating mode, mobilize several agents in parallel. This makes it possible to bypass certain bottlenecks and obtain more complete results without endlessly lengthening the wait.

‍

Multimodal dimension: text, image, video, audio

Muse Spark is designed from the outset as a natively multimodal model, which means it processes text, images and audio in an integrated way within a single architecture. Unlike systems that simply add vision or audio-processing modules on top of a language model, it leverages these modalities in a coherent manner. For example, it can combine an image description with an audio stream or a series of written questions to solve a complex task.

‍

A concrete example: Muse Spark can analyze an image, extract its key elements, then link them to external knowledge to produce a contextual answer, a product comparison or a technical analysis. This ability to “see and classify” lets it go beyond simple image description and enter the realm of visual reasoning, notably in sectors such as health, science or design. Although its outputs are still text-based for now, the flexibility of its multimodal input paves the way for richer responses in the future.

‍

Managing computational efficiency and accessibility

One of Meta's main bets with Muse Spark rests on its computational efficiency. According to the available information, the model reaches performance levels comparable to those of some large models from the previous generation, while using an order of magnitude less compute resources. This advance is made possible by a new architecture, extensive optimization of the pre-training stack and better data curation. The result: fewer FLOPs needed to reach a given level of skill.

‍

This reduced compute footprint has a direct impact on accessibility: deployment and inference costs go down, allowing more players (startups, SMEs, research teams) to take advantage of very large models. In addition, better efficiency in terms of tokens used for complex tasks makes large-scale deployments more sustainable, both economically and environmentally.

‍

In short, Muse Spark embodies a vision of AI that is not only more powerful, but also more accessible and more responsible.

‍

Muse Spark against the AI giants: OpenAI and Anthropic in its sights

Comparisons on global benchmarks

Muse Spark establishes itself as a serious competitor against the most advanced AI models. On the composite Artificial Analysis Intelligence Index, it scores 52, placing just behind GPT‑5.4 and Gemini 3.1 Pro, both around 57, and slightly behind Claude Opus 4.6 which reaches 53.

‍

This positioning places Muse Spark in the “frontier club” of the highest-performing general models, although it remains behind the leaders on extreme performance.

‍

When it comes to coding tasks, the gap becomes more pronounced. Muse Spark scores 59 on Terminal‑Bench 2.0, a respectable result but well behind GPT‑5.4 (close to 83) and Claude Opus 4.6 (around 81). In real-world agent task scenarios (GDPval‑AA), it reaches around 1,444 points, versus 1,674 for GPT‑5.4 and 1,607 for Claude Opus 4.6, confirming that OpenAI's and Anthropic's models still dominate complex autonomous workflows.

‍

Current strengths and limitations

Muse Spark's strengths lie mainly in its multimodal capability and in very specific areas. It posts impressive scores on MMMU Pro (around 80 to 81%), placing it on par with the best text-image models. Moreover, it excels on visual-understanding benchmarks such as CharXiv Reasoning, where it largely outperforms Claude Opus 4.6. On HealthBench Hard, it scores 42.8, slightly above GPT‑5.4 and Anthropic's models, reinforcing Meta's ambition to position Muse Spark as a valuable tool for health-related questions.

‍

On the other hand, Muse Spark shows weaknesses on abstract tasks and pure logic. On ARC‑AGI‑2, a benchmark designed to assess the recognition of new patterns, it reaches around 42.5, while GPT‑5.4 and Gemini 3.1 Pro exceed 76, nearly double. This reveals a fragility when facing very open-ended problems or those requiring fine-grained generalization.

‍

The “Contemplating mode” feature, which mobilizes several agents in parallel, notably improves its reasoning scores (HLE at 50.2%), but this advantage remains insufficient to close the overall gap on fully self-guided scenarios.

‍

Positioning within the Meta ecosystem

Muse Spark is not just a technical model; it represents a genuine strategic pivot within the Meta ecosystem. Natively integrated into Facebook, Instagram, WhatsApp and other services, it benefits from direct access to one of the largest social networks and databases of human interactions in the world.

‍

Unlike OpenAI's or Anthropic's models, often accessible through standard APIs, Muse Spark leverages knowledge of users' interactions, content, preferences and communication contexts to offer far more contextualized recommendations, assistants and services.

‍

This integration also allows Meta to precisely control the user experience, optimize costs and ensure consistency across features. Muse Spark thus positions itself as a personal AI engine, able to fit into browsing, gaming, shopping and communication without requiring an app switch. By opting for a free offering, while OpenAI and Anthropic reserve their highest-performing models for paid subscriptions, Meta is betting on mass adoption and loyalty within its ecosystem, thereby redefining the rules against the giants of general-purpose AI.

‍

Concrete applications and use cases

‍

For the general public: smart assistant and productivity

For individuals, Meta AI positions itself as an omnipresent smart assistant, integrated into WhatsApp, Messenger, Instagram and the new Meta AI app. It answers everyday questions instantly, helps organize the day, drafts messages, summarizes content or offers personalized recommendations thanks to the use of the Llama 4 model and web access.

‍

Available on mobile, desktop and soon via voice in its dedicated app, it offers a smooth experience for working, traveling, learning or having fun without having to switch interfaces.

‍

In everyday life, this assistant can, for example, summarize a blog article, plan a trip with transport and accommodation options, or help draft a resume or a professional email efficiently. With its built-in document editing and image generation features, it becomes a versatile tool for managing administrative tasks, creative projects or personal activities, while remaining accessible through apps already used by billions of people.

‍

For businesses and content creators

For businesses, Meta offers Business AI, a dedicated agent that analyzes a brand's social posts, advertising campaigns and website to provide personalized recommendations to customers, directly within chats or on the site. This solution aims to simplify AI adoption for small and medium-sized businesses by reducing integration costs and complexity.

‍

The Meta AI Business Assistant also helps advertisers optimize their campaigns via Ads Manager by offering suggestions, opportunity scores and support for resolving issues related to advertising accounts.

‍

Content creators, for their part, benefit from new generative tools for videos, images and music, integrated into Facebook and Instagram. These features make it possible to quickly produce advertising visuals, animations or music tailored to different audience segments, without requiring advanced technical skills.

‍

AI systems dedicated to creator discovery, collaboration recommendations and automated advertising partnerships streamline the matchmaking between brands and influencers. This improves the effectiveness of sponsored campaigns and organic content.

‍

Synergies with Meta platforms: Facebook, Instagram, WhatsApp

The strength of Meta AI lies in its tight integration with all of Meta's platforms. On WhatsApp, the assistant can reply to messages, summarize conversations and even generate images directly within chats. In addition, the handling of private data respects end-to-end encryption when the user enables the private processing options.

‍

On Instagram, Meta AI steps in within DMs, groups and certain feeds, making it easier to create content, find visual inspiration and interact with followers.

‍

On Facebook and Messenger, it enriches conversations, offers summaries of long content and helps keep groups up to date. Furthermore, the voice experiences in the Meta AI app make it possible to handle tasks hands-free while staying connected to your network. These synergies create an ecosystem where the assistant supports private exchanges, family communication and professional interactions, drawing on the preferences, profiles and content already present on the platforms to deliver personalized responses.

‍

Potential in health, science and connected devices

Beyond entertainment and marketing, the advanced versions of Meta AI and its Llama models open up prospects in demanding fields such as health and science. In the medical sector, AI can summarize complex records, explain diagnoses or protocols to non-specialist patients, or support research by quickly analyzing large volumes of scientific papers.

‍

Note-taking and lab-experiment structuring tools simplify the management of research projects and make collaboration between international teams easier.

‍

In the Internet of Things space, Meta AI finds its place in devices such as the Ray-Ban Meta glasses. These glasses can describe what the user sees, translate texts in real time or provide contextual information during visits, travel or work situations. These multimodal interactions, combining voice, vision and search, set the stage for more autonomous personal assistants, able to act as “copilots” in real life, whether for health monitoring or managing connected environments.

‍

What are the challenges and outlook for AI after Muse Spark?

Impacts on the balance of the AI market

The launch of Muse Spark marks a clear attempt by Meta to rebalance a market dominated by major US players such as OpenAI, Google and Anthropic. Thanks to a cutting-edge model integrated into its consumer apps, Meta establishes itself as a serious competitor, able to rival the leading AIs on technical benchmarks while capitalizing on its already massive user base.

‍

This technological advance could prompt other players to cut their prices or accelerate their update cycles, which would benefit businesses and users. Moreover, Meta's strategy of offering a free version with usage limits puts pressure on the paid business model currently dominant for most cutting-edge AIs.

‍

The rise of Muse Spark also reflects a broader trend: the concentration of technological resources in the hands of a few giants, while enabling a massive diffusion of advanced AI capabilities to a wide audience. This paradox shifts market dynamics, reinforcing the position of the major players while accelerating the democratization of generative AI tools.

‍

Ethical challenges and data sovereignty questions

The growing centralization of AI in the hands of large platforms raises major concerns around privacy and data sovereignty. Like its competitors, Muse Spark relies on massively collected data, often drawn from public or semi-public interactions. This raises questions about consent procedures and about users' control over their own information.

‍

Regulators are closely examining the use of personal data to train these models, particularly in sensitive areas such as health. Although Meta states that it works with healthcare professionals to ensure the reliability of results, these practices remain subject to debate.

‍

In Europe and elsewhere, the rise of such models highlights the urgent need for specific legislation covering the transparency of AI systems, accountability for their decisions and the protection of fundamental rights. Users, for their part, will need to learn to distinguish legitimate uses of generative AI from those that could threaten their privacy or their decision-making autonomy.

‍

Toward a new era of accessible generative AI

With Muse Spark, Meta is accelerating the transition toward generative AI that is accessible and embedded in the daily lives of millions of people. The advanced reasoning, multimodal processing and agent orchestration capabilities now enable the automation of complex tasks once reserved for experts.

‍

This evolution paves the way for an era in which AI becomes an omnipresent assistant, supporting creativity, learning, project management and even decision-making in certain fields.

‍

Looking ahead to 2026 and beyond, the rise of models such as Muse Spark could normalize the use of generative AI in professional and personal spheres. However, this transformation will depend on the ability to address the associated technical, ethical and regulatory challenges. If these obstacles are overcome, this new phase could redefine the way individuals interact with technology, turning AI into a collaborative tool that is accessible, rather than an opaque device reserved for an elite.

‍

Conclusion

All in all, ignoring the payment of a citation is not a consequence-free option. That said, it is understandable that not everyone can pay immediately. The good news is that there is a simple, fast and free alternative: requesting a 100% reduction of your citation amount online.

‍

The procedure is extremely easy: it takes only three minutes, requires no travel and is done simply by filling out a form. This request may be accepted or refused, but it is completely free. Why not give it a try? It costs you nothing.

‍

So don't wait any longer: open your browser, enter your case number and submit your reduction request. The longer you delay, the more your situation could become complicated. In just a few minutes, you can check whether you are eligible for a full refund of your citation. You have nothing to lose and everything to gain.

‍

FAQ

‍

What is Muse Spark and what are its main features?

Muse Spark is the brand-new proprietary, multimodal AI model developed by Meta Superintelligence Labs. It supports text, images, audio and tools, and offers three distinct modes: Instant (fast), Thinking (chain of thought) and Contemplating (parallel multi-agent). This model is available for free on meta.ai and stands out for its performance in fields such as science, mathematics and health.

‍

How does Muse Spark compare to models like GPT-5 or Claude Opus 4.6?

Muse Spark is designed to compete with models such as GPT-5.4 and Claude Opus 4.6. It stands out for its token-management efficiency, its multimodal compatibility (notably vision) and its creativity. For example, it achieves an impressive score of 42.8% on the HealthBench Hard benchmark. However, it remains slightly behind in certain areas such as coding (77.4% on SWE-bench versus 80.8% for its competitors), agentic tasks and the GDPval-AA evaluation. Claude and GPT remain the global leaders in these areas.

‍

Is Muse Spark free and how do you access it?

Yes, Muse Spark is entirely free for the general public. You can access it by going to meta.ai or via the Meta AI app, after creating a free account. The model is also integrated into Meta's mobile apps for greater accessibility.

‍

A paid API access is planned for the future, although the pricing details have not yet been communicated.

‍

What are Muse Spark's multimodal capabilities, especially in health?

Muse Spark is a natively multimodal model that supports text, images, audio and tools. It excels at visual reasoning and in tasks involving multiple agents. When it comes to health, it stands out particularly thanks to a database enriched by the work of 1,000 doctors. It reaches a notable score of 42.8% on HealthBench Hard, demonstrating its expertise in nutrition, medications and physiology. It also offers interactive displays for a better understanding of medical information.

‍