Meta launches Muse Spark frontier model and pivots from open-source to proprietary AI

Meta has officially re-entered the top tier of the artificial intelligence race with the release of Muse Spark, a frontier model that represents a total architectural and strategic departure for the social media giant.[1][2][3] Developed by the newly formed Meta Superintelligence Labs, Muse Spark is the company’s first high-end model to be released without open weights, a move that effectively ends the era of the Llama-driven open-source playbook that Meta championed for years.[4] The launch follows a nine-month period of internal restructuring and a reported 14.3 billion dollar investment into the AI data-labeling firm Scale AI, which culminated in the hiring of Alexandr Wang as Meta’s inaugural Chief AI Officer.[5] This release signals Mark Zuckerberg’s intent to move past the perceived performance stumbles of previous models and compete directly with the proprietary systems of OpenAI, Google, and Anthropic.

The technical foundation of Muse Spark is a complete ground-up rebuild of Meta’s pretraining stack.[4] According to technical documentation released alongside the model, the architecture is natively multimodal, meaning it was trained to process text, image, audio, and video data within a single unified framework rather than relying on disparate encoders. A defining feature of the model is its new Contemplating mode, a reasoning system that differentiates itself from the serial chain-of-thought methods used by competitors.[5][3][6] Instead of thinking in a linear sequence, Muse Spark orchestrates multiple sub-agents in parallel to solve complex problems.[3][6][7] This approach is paired with a training technique Meta calls thought compression, which penalizes the model during reinforcement learning for excessive reasoning tokens.[5] The company claims this allows Muse Spark to achieve frontier-level reasoning while using an order of magnitude less compute than the midsize variants of its predecessor, Llama 4.[5][8]

Independent testing highlights a model that has successfully closed the competitive gap, though it does not yet claim the undisputed top spot in the industry.[2] On the Vals Index, a respected third-party benchmark for domain-specific tasks, Muse Spark debuted in third place with a score of 65.66 percent, trailing only Anthropic’s Claude Sonnet 4.6 and Claude Opus 4.6. In terms of scientific reasoning, the model scored 89.5 percent on the GPQA Diamond benchmark, placing it just behind Google’s Gemini 3.1 Pro and OpenAI’s GPT-5.[5]4. However, the model showed significant leadership in specialized verticals, particularly in healthcare.[5][7] By collaborating with over 1,000 physicians to curate its training data, Meta enabled Muse Spark to score 42.8 percent on the HealthBench Hard evaluation, vastly outperforming many of its peers.[5] Conversely, the model continues to lag in abstract reasoning tasks, scoring 42.5 on the ARC AGI 2 benchmark compared to the mid-70s range seen in the top-tier models from Google and OpenAI.[5]

The decision to keep Muse Spark’s weights closed is perhaps the most significant policy shift in the AI industry this year. For years, Meta was the primary benefactor of the open-source community, providing the weights for the Llama series and allowing developers to build a massive ecosystem of fine-tuned versions. By pivoting to a proprietary, API-driven model, Meta is prioritizing product control and monetization over community adoption. Industry analysts suggest this shift was inevitable as the cost of developing frontier models began to reach the hundred-billion-dollar range, requiring a clearer path to revenue than open releases could provide. Meta has framed the closed nature of Muse Spark as a temporary necessity for safety and competitive reasons, noting that while this initial version is proprietary, the company maintains a scaling ladder that may include open-source releases for smaller or future generations of the Muse family.

Within the Meta ecosystem, Muse Spark is being positioned as the engine for personal superintelligence.[9][3][10][11] It is already powering the Meta AI assistant on the web and through dedicated mobile apps, with a broader rollout scheduled for WhatsApp, Instagram, and Facebook in the coming weeks.[10][3] The model’s multimodal strengths are being leveraged for new consumer features, such as a shopping mode that can analyze a user’s social media feed to offer personalized styling recommendations. It is also designed to integrate deeply with Meta’s Ray-Ban smart glasses, where its ability to process real-time visual and audio data allows it to act as an embodied assistant capable of recognizing environments and providing localized information. For enterprise users, Meta has launched a private API preview for select partners, marking the beginning of a more traditional cloud-based AI business model.

The implications for the broader AI industry are profound, as Meta’s pivot leaves a void in the high-end open-weight market that few other companies have the capital to fill. While players like Mistral and xAI continue to release capable open models, none have yet reached the frontier performance level that Muse Spark currently demonstrates. The shift also highlights a growing trend toward inference-time compute and agentic orchestration as the next battleground for AI leadership. By focusing on multi-agent parallel reasoning, Meta is betting that the path to artificial general intelligence lies in efficiency and the ability of models to plan and execute tasks across different data formats.

As Meta scales its new Muse series, the focus will likely turn to how the company balances its historical commitment to openness with the commercial demands of being a frontier AI developer.[12] The successful launch of Muse Spark has already had a positive impact on the company’s market valuation, with shares rising nearly 7 percent following the announcement.[2] For now, Meta has proven it can still compete at the highest levels of model performance. Whether it can maintain this momentum while locked into a more closed development cycle will determine if Zuckerberg’s massive capital expenditures in superintelligence labs can truly disrupt the dominance of his rivals in the years to ahead. The arrival of Muse Spark marks the end of Meta's period of relative silence in the LLM space and begins a new, more aggressive chapter in its quest for AI supremacy.