Meta has taken a significant step in its artificial intelligence strategy with the launch of Muse Spark , a new model that marks a turning point following the restructuring of its AI division. Mark Zuckerberg's company is attempting to regain ground against rivals like OpenAI, Google, and Anthropic with a more integrated approach across its products and a clearly global ambition.
This model will power Meta AI on the web, the standalone app, and, progressively, in services like WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban Meta AI glasses . Although the rollout has begun in the United States, the roadmap includes international expansion that is expected to affect users in Spain and the rest of Europe as soon as the region's regulatory and privacy issues are resolved.
A strategic shift: from Llama to the Muse family
Muse Spark is the first model in the new Muse series developed by Meta Superintelligence Labs (MSL) , the elite laboratory Zuckerberg established after investing billions in talent and infrastructure. At its helm is Alexandr Wang, founder of Scale AI, whose appointment was finalized following an investment of approximately $14.300 billion (around €12.250 billion) in his former company.
The model was developed over approximately nine months, and during this time, Meta claims to have rebuilt its AI stack virtually from scratch . The result is a system that doesn't aim to be the largest on the market, but rather one of the most efficient: small, fast, and capable of handling complex queries without requiring excessive computing power, improving upon the performance of previous versions like Llama 4 or Llama 4 Maverick.
This change of direction comes after the lukewarm reception of the Llama family, which, despite its open-source approach, failed to position Meta at the level of OpenAI, Anthropic, or Google's Gemini . With Muse Spark, the company acknowledges the need for a product better suited to its real-world applications while also being capable of sustaining its own business model around AI.
The markets have responded positively. Following the announcement of Muse Spark, Meta's shares rose between 6% and 8% on Wall Street , a sign that investors interpret this move as a serious attempt by the company to continue competing in the next major technology platform.
Proprietary model, but without closing the door to open source.
One of the most striking points is that Muse Spark is a closed, proprietary model . This contrasts with Meta's previous approach of releasing Llama weights and models under relatively open licenses. Now, the company prioritizes control over the technology and its commercial exploitation, although it maintains that future versions of the Muse family could offer open variants.
According to Meta, improvements in the training process and infrastructure allow a smaller model to achieve performance comparable to mid-range versions of Llama 4 , with an order of magnitude less computing power. This efficiency is key to scaling the use of Muse Spark across all applications without significantly increasing costs.
In internal and third-party assessments, the company indicates that the model offers competitive performance in multimodal perception, reasoning, and agentic tasks , although it admits that there is still room for improvement in areas such as complex programming flows or very long-range agent systems.
External experts have pointed out that the lack of open weights will make it difficult for the community to accurately measure the model's technical value. Even so, Meta emphasizes that it will continue to invest in open projects in parallel , so that its AI ecosystem is distributed among internal models, commercial products, and community developments.
Key capabilities: multimodality, health, and procurement
In practical terms, Muse Spark presents itself as a multimodal AI capable of working with text and images . It can analyze photographs, understand their content, and combine them with written instructions—a particularly relevant area for devices like Ray-Ban Meta AI glasses or for the enormous volume of images and videos circulating on the company's social media.
Meta places a strong emphasis on health. Muse Spark has been trained using curated data with the help of over 1.000 doctors, who have reviewed and refined information on nutrition, exercise, and other health topics . The goal is for the assistant to be able to answer medical questions more accurately, generate visual explanations, and help understand complex concepts, such as how certain muscles function during exercise or the nutritional composition of a dish.
This approach is not without risks, especially in Europe, where data protection criteria and regulatory requirements for digital health services are particularly stringent . In any case, Meta claims to have subjected the model to boundary risk assessments, alignment tests, and adversarial robustness tests, with results that place Muse Spark within the safety margins assumed by the industry.
Another key element of the offering is Shopping mode . Muse Spark connects with Instagram, Facebook, and Threads to suggest products, styles, or purchase recommendations based on the creators and communities the user already follows. When someone asks about clothing, home decor, or gift ideas, the AI ​​can combine that social information with the context of the query to display potentially more relevant options than a generic search engine.
In more everyday scenarios, such as planning a trip, organizing an itinerary, or comparing destinations for families with children , the model distributes the task among several AI agents: one can put together the agenda, another check accommodations, another look for specific activities, and then combine everything into a coherent proposal for the user.
Reasoning modes and agent-based architecture
Like other next-generation models, Muse Spark offers several reasoning modes to suit each type of query. Meta describes three main levels: a quick mode, geared towards instant answers; an intermediate mode for tasks that require more elaboration; and an advanced mode, designed for in-depth analysis.
In practice, these levels translate into what the company calls modes such as "Instant," "Thinking," and "Contemplating ." The first prioritizes speed for simple questions; the second focuses on more complex problems, such as summarizing documents, analyzing contracts, or explaining scientific concepts; and the third activates a set of agents that work in parallel to obtain a more elaborate answer, in line with the so-called extreme reasoning modes present in models like Gemini Deep Think or certain advanced GPT plans.
To make this approach viable without skyrocketing costs, Meta's engineers have introduced a penalty on thinking time during training. In other words, they push the model to compress its reasoning and use fewer tokens when faced with complex problems, trying to maintain response quality while reducing latency and resource consumption.
The company states that, in benchmark tests, Muse Spark performs competitively against models like Claude Opus 4.6 or advanced versions of GPT in areas such as multimodal reasoning and certain agent tests. However, it also acknowledges weaknesses in areas like sophisticated programming and very long-running agent tasks, where other packages still have an advantage.
In any case, the parallel agent architecture and the ability to switch between response modes suggest that Meta wants its assistant to move beyond being a simple chatbot and become a "personal superintelligence ," capable of executing specific steps, coordinating subtasks, and moving with some autonomy within the company's ecosystem.
Full integration into Meta apps and its potential impact in Europe
Muse Spark is already behind Meta AI in the web version and dedicated app in the United States, and the company has confirmed that in the coming weeks it will be rolled out to WhatsApp, Instagram, Facebook, and Messenger, as well as to Ray-Ban Meta AI glasses. The idea is that users won't have to install anything extra: the AI ​​will be integrated as a layer across the services that millions of people use daily.
In Europe, and specifically in Spain, the arrival of these features will be contingent upon regulations such as the General Data Protection Regulation (GDPR) and the new European regulation on AI . Muse Spark's ability to leverage text, images, videos, and user reactions raises questions about how consent, profiling, and the use of personal data to hyper-personalize responses will be managed.
Meta sees this abundance of information as one of its major competitive advantages. By having access to social interactions, audiovisual content, and explicit reactions (likes, comments, shares) , it argues that it can train and refine models that better understand people's preferences. In Europe, however, it will have to balance this potential with strict requirements regarding transparency, data minimization, and the possibility of opt-out.
Beyond social media, the deployment of Muse Spark in products like smart glasses opens the door to "ambient computing" experiences : pointing the camera at an object, receiving instant contextual information, or requesting nutritional guidance without having to manually scan labels. These types of uses, again, will have to be closely scrutinized within the European regulatory framework.
If Meta manages to adapt Muse Spark to these legal frameworks without significantly diminishing its capabilities, European users could benefit from more proactive and contextual assistants , both in content consumption and in shopping, entertainment, and wellness. The company, for its part, would gain a stronger position compared to the models already operating in the region through browsers and external applications.
Business, payment APIs, and massive investment in infrastructure
In addition to direct use in its applications, Meta is testing restricted API access to Muse Spark for select partners . The idea is to gradually roll out a paid service that allows companies and developers to integrate the model into their own products, generating a new revenue stream beyond advertising.
Until now, the company had not managed to establish a significant business based on selling AI capabilities as a service , unlike other players in the sector. With a proprietary model focused on very specific tasks, it hopes to position itself as an alternative in a market already dominated by platforms backed by Microsoft, Google, and other cloud providers.
This move is accompanied by a very aggressive increase in investment in data centers, chips, and networks . Meta has announced that its AI-related capital expenditure for 2026 will be extremely high, nearly double the previous year's figure, with the goal of training, deploying, and running increasingly complex models for hundreds of millions of users.
The positive stock market reaction suggests that, at least for now, markets are willing to tolerate this increase in capital expenditure in exchange for the possibility that the company will secure a significant role in the generative AI economy . However, the real test will come when the adoption rate of its API becomes clear, along with the willingness of European and American companies to build upon a technology whose public roadmap is still limited.
Internally, Meta claims to have maintained a relatively flat structure and autonomy for its research teams , with Wang directly overseeing around 100 people. The intention is to move with agility in an environment where development cycles for leading-edge models are shortening significantly.
The launch of Muse Spark reflects a shift in Meta's AI strategy: from one focused on open models that failed to gain traction in the market, to a more pragmatic, proprietary approach geared towards specific products . If the model delivers on its promises in terms of performance, integration, and security, the company could close the gap with current leaders; if it falls short, the opportunity cost of this reorientation will be significant in a rapidly evolving sector.