Meta launches proprietary Muse Spark AI model from new Superintelligence Labs

Meta has launched Muse Spark, its first AI model from the newly formed Meta Superintelligence Labs, marking a significant shift in the company's strategy after a $14.3 billion investment in Scale AI and a ground-up overhaul of its AI infrastructure. Led by former Scale AI CEO Alexandr Wang, the model was rebuilt from scratch over nine months and is now powering queries on the Meta AI app and Meta.ai website in the US, with rollout planned for WhatsApp, Instagram, Facebook, and Messenger in the coming weeks.^[1][2][3]

Unlike Meta's previous Llama family of mostly open-source models, Muse Spark is initially closed-source and proprietary, accepting voice, text, and image inputs while producing text-only outputs.^[4][6] According to Bloomberg, this debut comes as CEO Mark Zuckerberg pushes to compete with rivals like OpenAI, Anthropic, and Google amid a multibillion-dollar AI reorganization.^[2] TechCrunch describes it as the first release under Superintelligence Labs, emphasizing a "contemplating" reasoning mode that runs sub-agents in parallel for enhanced performance.^[3][1]

Benchmark results position Muse Spark as competitive but not dominant among top models. It scores 52 on the Artificial Analysis Intelligence Index, ranking in the top 5 and ahead of models like Claude Sonnet 4.6 and Grok 4.20, though behind Gemini 3.1 Pro Preview, GPT-5.4, and Claude Opus 4.6.^[1] On GPQA Diamond, a PhD-level reasoning test, it achieved 89.5%, trailing Gemini 3.1 Pro's 94.3% and scores from Claude Opus 4.6 and GPT-5.4 around 92-93%, as reported by Fortune.^[2] The model excelled on HealthBench Hard with 42.8%, outperforming rivals including Opus 4.6 and Gemini 3.1 Pro, and showed token efficiency using just 58 million output tokens for key evaluations—comparable to leaders but far less than some competitors.^[1][2]

Meta acknowledges gaps in agentic tasks and coding, where Muse Spark trails models like GPT-5.4 and Claude Sonnet 4.6 on benchmarks such as GDPval-AA and TerminalBench Hard.^[1] In its technical blog, the company highlights improvements in architecture, optimization, and data curation, claiming over an order of magnitude less compute than predecessor Llama 4 Maverick, with a new reinforcement learning pipeline for predictable gains.^[2][5] Ars Technica notes strong results in multimodal perception, reasoning, health, and some agentic tasks, but ongoing investments target weaknesses.

This launch revives Meta's standing after Llama 4's poor reception in April 2025, as VentureBeat points out, potentially signaling a move away from open-source dominance toward proprietary advancements.^[4] The Verge reports immediate integration into Meta's ecosystem, affecting billions of users across its platforms and raising questions about accessibility versus performance. Slashdot mentions plans for an open-source version later, which could balance commercial goals with Meta's research heritage.^[6]

For users and developers, Muse Spark means more capable AI interactions in everyday apps, from voice queries to image analysis, though text-only outputs limit some creative uses initially. Businesses and researchers watching the AI race will scrutinize independent verifications of these benchmarks, as Hacker News discussions already question real-world programming utility compared to leaders like Opus.^[5][6]

Looking ahead, Meta positions Muse Spark as the first rung on a "scaling ladder" toward personal superintelligence, with each iteration building on validated progress before larger models.^[2][5] As the company addresses admitted shortcomings in long-horizon agents and coding, this could reshape competition, influencing how AI integrates into social media, messaging, and beyond for Meta's vast audience.