Meta has officially introduced the Meta Muse Spark AI model, a new natively multimodal artificial intelligence system that marks a sharp departure from the company’s open-source legacy. Developed over nine months by the newly formed Meta Superintelligence Labs, the model aims to directly challenge industry leaders like OpenAI, Google, and Anthropic. The platform is already powering the Meta AI assistant on the web and mobile app, signaling a massive, multibillion-dollar overhaul of the tech giant’s AI ambitions.
The launch of the Meta Muse Spark AI model, internally codenamed Avocado, represents a fundamental shift in strategy. Unlike the open-source Llama series that defined the company’s previous artificial intelligence efforts, this new iteration is a closed, proprietary system. Meta rebuilt its entire artificial intelligence infrastructure, architecture, and data pipelines from scratch to create it. The model can process text, images, and audio, though it currently only outputs text.
Advanced Reasoning and Multi-Agent Capabilities
The system introduces three distinct reasoning modes tailored for different tasks. The Instant mode handles casual queries quickly, while the Thinking mode uses a chain-of-thought process for harder questions. The standout feature is the Contemplating mode, which tackles complex problems by deploying multiple sub-agents to work in parallel.
Meta claims this parallel processing allows the system to achieve advanced reasoning while using ten times less computing power than its predecessor, Llama 4 Maverick. The company achieved this efficiency through thought compression, a training technique that penalizes the model for excessive thinking time. This forces it to solve problems with fewer steps without losing accuracy.
A Mixed Bag on Performance Benchmarks
Following the announcement, Meta’s shares rose six percent according to Futurism, while CryptoRank and The Next Web reported a climb of as much as nine percent. Despite this investor enthusiasm, performance benchmarks present a complicated picture. The model currently ranks fourth on the Artificial Intelligence Index v4.0, trailing behind Google’s Gemini 3.1 Pro Preview, OpenAI’s GPT-5.4, and Anthropic’s Claude Opus 4.6.
The software struggles with abstract reasoning and coding tasks compared to its rivals, but it excels in highly specific areas. On the CharXiv Reasoning benchmark, which tests visual understanding of charts and figures, the platform scored 86.4, beating both Gemini 3.1 Pro and GPT-5.4.
The system also shines in the medical field. Meta collaborated with over 1,000 healthcare professionals to curate training data, helping the model score 42.8 percent on the HealthBench Hard evaluation. This easily outpaced GPT-5.4 and Gemini 3.1 Pro, positioning the tool as a powerful resource for nutrition, dietary health, and medical reasoning.
Massive Investments and Infrastructure Upgrades
The development follows a turbulent period for Meta’s technology division. After dissatisfaction with the progress of previous models and controversies surrounding allegedly faked benchmark results for Llama 4—which reportedly led former artificial intelligence head Yann LeCun to leave the company—CEO Mark Zuckerberg initiated a massive structural overhaul. Meta abandoned its two-trillion-parameter Llama 4 variant, codenamed Behemoth, and shifted its focus to Meta Superintelligence Labs.
To lead this new division, Meta hired Alexandr Wang as its first-ever chief AI officer. The company also secured a 49 percent non-voting stake in Wang’s firm, Scale AI, for $14.3 billion.
Financial commitments do not stop there. Meta recently expanded its cloud partnership with CoreWeave, signing a $21 billion, multi-year agreement for dedicated computing capacity running through 2032 on Nvidia’s Vera Rubin platform. Furthermore, the company expects to spend between $115 billion and $135 billion on capital expenditures in 2026 alone, while pledging $600 billion toward U.S. infrastructure through 2028. These massive investments follow an $80 billion loss in Meta’s Reality Labs division, which prompted widespread job cuts.
Privacy Concerns and Rollout Plans
As the software rolls out, it brings potential privacy and accessibility concerns. Users must log in with an existing Meta account, such as Facebook or Instagram, to access the tool. Because Meta intends to build a personal superintelligence, this tight integration raises questions about how user data will be utilized. Additionally, while the model is currently free, Meta executives have indicated they are considering putting future capabilities behind a paywall. The development methods have also faced scrutiny, specifically the controversial use of third-party open-source models, including one from Alibaba.
Currently available in the United States, the platform will soon expand across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban smart glasses. Meta is also offering a selective private API preview to chosen partners. While the company hopes to open-source future versions, the current release serves as a clear signal that Meta is willing to lock down its technology to secure a dominant position in the industry.
