Meta launches Muse Spark and returns to the AI ​​competition

Muse Spark it’s the new one artificial intelligence model announced by Meta and is the first developed by Meta Superintelligence Labsthe division created by the group to strengthen its position in the global race forAI. The model was presented on Wednesday and is the first major release after joining the company Alexander Wangformer CEO and co-founder of Scale AI, who has held the role of Chief AI Officer of Meta since June (the first in the group’s history) as part of the $14.3 billion operation with which Meta acquired a 49% non-voting stake in the company. Muse Spark opens a new family of models, the “Muse” series, and was developed in-house under the code name “Avocado”.

According to Meta, the system was built in about nine months of work, during which time the company rebuilt its own AI infrastructure and the so-called “stack”, i.e. the set of models, systems, training pipelines and operational tools necessary to develop and deploy artificial intelligence in the group’s products.

The launch comes after a complicated phase for the Meta’s AI strategy. The models open source most recent of the Llama 4 family, presented last April, did not have the desired impact on developers and the market. Hence the group’s decision to reorganize the sector, accelerate investments and focus on a new dedicated division. Meta did not want to present Muse Spark as the largest or most advanced model ever, preferring to focus on a “small and fast” reality, designed to be efficient and to offer competitive performance in various activities. In a technical post, the company describes it as a multimodal model(capable of accepting voice, textual and visual input, while producing text-only output for now) with reasoning abilityuse of external tools and multi-agent orchestration.

It is also the first reasoning model by Meta: unlike the previous ones, which produced immediate responses based on training, Muse Spark is able to tackle problems progressively, changing strategy if the initial approach doesn’t work. One of the central elements of the launch is thecomputational efficiency.

Meta claims that the new system, thanks to revised training techniques and a new infrastructure, is capable of achieving capabilities comparable to its previous mid-range variant while using an order of magnitude less computing power. This aspect is relevant because reducing the calculation requirement means lowering costs, times and demand for infrastructure for large-scale use.

Applications and integrations

On the level of applicationsMuse Spark already powers the assistant Meta AI in the dedicated app and on the group’s desktop site. The company of Zuckerberg he also indicated that the model will also be extended in the coming weeks to Facebook, Instagram, WhatsApp, Messenger and glasses Meta-Ban Ray. Looking ahead, the model will also have to support other functions, including Vibes AI, the video product of the Meta AI ecosystem, which currently relies on third-party models such as Black Forest Labs.

Among them functionality announced there are different methods depending on the complexity of the requests. Users will be able to switch between a quicker mode for simple questions and a more advanced one for more complex searches (queries).including the analysis of documents and images. The system can also launch multiple subagents in parallel to address more complex requests. For example, planning a trip with one agent writing the itinerary, another comparing destinations, and a third looking for specific activities, all at the same time. One is also planned shopping modedesigned to suggest purchases of clothing and home solutions, combining the linguistic model with data on user interests and behavior.

Monetization and distribution strategy

The group also opened a first channel monetization. Muse Spark is offered, in this initial phase, via API private to a limited number of selected partners, with the intention of expanding access to a wider audience of developers at a later stage. This is a significant step, because it broadens the role ofAI in the business model of the group.

Another important element concerns the distribution strategy. Muse Spark is a proprietary modelin clear discontinuity with the approach open source followed with the Llama family. However, Zuckerberg stated publicly on Threads that the group plans to release “increasingly advanced open source models” in the future, leaving open the possibility of open distribution for subsequent releases.

Future performance and investments

On the front of performanceMuse Spark lags behind market leaders in some areas (particularly coding) but is competitive in other segments, including multimodal perception and health related questions. The launch is part of a phase of strong increase in spending on artificial intelligence from Meta.

AI-related capital expenditures in 2026 will be between $115 billion and $135 billion, nearly double the previous year. Meta aims to use Muse Spark as a base to strengthen presence in consumer productstest new API revenue and catch up compared to OpenAI, Anthropic e Googlewhich today dominate the advanced model segment.