Meta Launches Muse Spark, Its First Closed-Source AI Model, as Superintelligence Labs Bets on Proprietary Multimodal Reasoning
Meta releases Muse Spark, a proprietary multimodal reasoning model from its new Superintelligence Labs unit, marking a sharp departure from its open-weight Llama strategy after a nine-month infrastructure rebuild.
Overview
Meta on April 8 released Muse Spark, the first model produced by its Meta Superintelligence Labs division, in what amounts to the company’s sharpest strategic turn in the AI race to date. Unlike every entry in the Llama family that preceded it, Muse Spark is proprietary: no open weights, no downloadable checkpoints, and access limited to Meta’s own products and a private API preview for select partners, according to TechCrunch. The move follows a nine-month infrastructure rebuild led by Alexandr Wang, the former Scale AI co-founder whom Meta hired as its first chief AI officer after investing $14.3 billion for a 49 percent stake in Scale AI, as Fortune reported.
The release comes after Meta previously delayed its next-generation model, codenamed Avocado, amid internal benchmark shortfalls against Google, OpenAI, and Anthropic.
What We Know
Muse Spark is described by Meta as a natively multimodal reasoning model with support for tool use, visual chain of thought, and multi-agent orchestration, according to Meta’s technical blog. The company says it rebuilt its entire pretraining stack over nine months, overhauling model architecture, optimization, and data curation.
A key technical feature is Contemplating mode, which orchestrates multiple agents reasoning in parallel rather than relying on extended single-agent inference. Meta reports that this mode achieves 58 percent on Humanity’s Last Exam and 38 percent on FrontierScience Research, according to Meta’s technical blog. The company claims the model reaches capabilities equivalent to its earlier Llama 4 Maverick with over an order of magnitude less compute.
Meta’s own benchmarks, reported by Fortune, show Muse Spark competitive but not dominant against frontier rivals. On GPQA Diamond, a PhD-level reasoning benchmark, Muse Spark scored 89.5 percent, trailing Gemini 3.1 Pro Preview at 94.3 percent, GPT-5.4 at 92.8 percent, and Claude Opus 4.6 at 92.7 percent. The model performed strongest on HealthBench Hard, where it led with 42.8 percent. Meta acknowledged “performance gaps, specifically long-horizon agentic systems and coding workflows,” according to Fortune.
The model is currently available through the Meta AI app and meta.ai website, with a rollout to WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban AI glasses planned in the coming weeks, according to Meta’s announcement. Users must log in with existing Meta accounts to access it.
The Open-Source Question
The proprietary release represents a conspicuous break from the strategy that defined Meta’s AI positioning for years. The company built significant developer goodwill through the Llama model family, which was released with open weights and attracted a large ecosystem of fine-tuners, researchers, and commercial adopters. With Muse Spark, few people outside Meta’s product ecosystem can use the new model, as Fortune noted.
Wang addressed the shift, stating that bigger models are already in development “with plans to open-source future versions,” according to TechCrunch. Meta’s announcement similarly references planned future open-source models, though it offers no timeline or specifics.
What We Don’t Know
Meta has not published model size, architecture details, or training dataset composition for Muse Spark. The company has disclosed no pricing for the private API preview, and the scope of partner access remains undefined. Whether the promised open-source follow-ups will match Muse Spark’s capabilities or constitute smaller, less capable variants is unclear.
The broader financial picture also remains uncertain. Meta has committed hundreds of billions of dollars to AI computing infrastructure and offered AI researchers pay packages running into the hundreds of millions of dollars including equity, according to Fortune. Whether Muse Spark can generate revenue proportional to that investment, particularly without an external developer platform, is an open question.
It is also worth noting that Llama 4, released in April 2025, was widely criticized after Meta later admitted using specialized, unreleased model versions fine-tuned for specific tasks to inflate benchmark scores, as Fortune reported. Whether the Muse Spark benchmarks reflect the production model’s actual performance has not been independently verified at the time of launch.
Analysis
Muse Spark’s closed-source release signals that Meta’s leadership has concluded open weights alone cannot sustain a competitive frontier AI business. The company’s three-billion-user distribution advantage across Facebook, Instagram, WhatsApp, and Messenger gives it a deployment channel that no other AI lab can match, but only if the model stays proprietary and tied to those surfaces. Meta’s stock rose 9 percent on the announcement, suggesting investors approve of the pivot toward a more defensible commercial model.
The tension between this approach and Meta’s cultivated open-source reputation will likely define the next phase of the AI landscape. If Muse Spark proves competitive enough to retain users within Meta’s ecosystem, the incentive to release open-weight successors diminishes significantly. For the developer community that built around Llama, the practical question is whether Wang’s promise of future open-source releases materializes, or whether Muse Spark marks the beginning of a permanent strategic realignment.