Meta’s Muse Spark
Meta launched Muse Spark, a new closed-source large language model positioned to compete on performance and control rather than openness. (thenextweb.com). Coverage notes the model leads on some health benchmarks but still trails on abstract reasoning, reflecting Meta’s strategy to trade openness for product differentiation. (thenextweb.com).
Large language models are prediction engines: they take a pile of text, images, and examples, then guess the next useful token the way your phone guesses the next word in a sentence. Meta just put a new one called Muse Spark at the center of WhatsApp, Instagram, Facebook, Messenger, and its artificial intelligence glasses instead of releasing it the way it released earlier Llama models. (about.fb.com) For three years, Meta’s artificial intelligence pitch was openness. The Llama family was released with model weights that outside developers could download, fine-tune, and run on their own hardware, which helped Meta build mindshare against OpenAI and Google. (wired.com) Muse Spark breaks from that playbook. Meta says the model is closed, built by Meta Superintelligence Labs after a nine-month rebuild, and available through Meta’s own products rather than as downloadable weights. (about.fb.com) (thenextweb.com) Meta is selling the switch as control plus speed. The company says Muse Spark is “purpose-built to prioritize people,” powers a faster Meta AI assistant, and is the first model in a new Muse family aimed at what Mark Zuckerberg calls “personal superintelligence.” (about.fb.com) (techcrunch.com) The model is also trying to do more than plain text chat. Meta describes Muse Spark as natively multimodal, which means it is built to handle text and images together from the start, with tool use, visual chain of thought, and multiple software agents working in parallel on harder tasks. (about.fb.com) (siliconangle.com) The benchmark story is mixed in a very specific way. The Next Web reports Muse Spark leads on some health tests, while still trailing on abstract reasoning, which suggests Meta tuned it for concrete product use cases before trying to win every general leaderboard. (thenextweb.com) Independent testing points the same way. Artificial Analysis says Muse Spark scores 52 on its Intelligence Index, behind Gemini 3.1 Pro at 57, GPT-5.4 at 57, and Claude Opus 4.6 at 53, even as Meta’s model posts standout results in some narrower categories. (artificialanalysis.ai) That tradeoff fits Meta’s business better than the old open model strategy did. If the best version of Meta’s model only lives inside Meta apps, then every improvement helps Meta keep users inside its own products instead of handing the same advantage to every rival developer for free. (wired.com) (aol.com) The timing also says a lot about where Meta thought it stood. TechCrunch reports Meta created Meta Superintelligence Labs after Zuckerberg grew unhappy with how Meta and Llama were lagging behind ChatGPT and Claude, and Muse Spark is the first public output from that reset. (techcrunch.com) So Muse Spark is not Meta trying to win the old open-source argument one more time. It is Meta deciding that the next fight is over distribution, product control, and keeping the smartest model features inside the apps used by billions of people every month. (about.fb.com) (thenextweb.com)
