Meta launches Muse Spark

Meta just put its newest artificial intelligence model inside its own apps first, not out on the open web for everyone. Muse Spark went live on April 8 in the Meta AI app and website, with Meta saying WhatsApp, Instagram, Facebook, Messenger, and its artificial intelligence glasses will get it in the coming weeks. (about.fb.com) That is a change in strategy from the company that spent years promoting its Llama models as open-weight software other developers could download and build on. Meta now says Muse Spark will be offered only in private preview through an application programming interface, which is the rented pipe developers use to send prompts to a company’s model. (cnbc.com) (about.fb.com) The pitch is not “biggest model wins.” Meta says Muse Spark is “small and fast by design,” which means it is built more like a compact engine that can answer quickly inside consumer apps than like a giant lab demo that burns through expensive computing chips on every request. (about.fb.com) (cnbc.com) Meta also says the model is multimodal, which means it can take in more than one kind of input, including text, voice, and images. In Meta’s examples, a user can point a camera at an airport snack shelf and ask which items have the most protein, instead of typing every label by hand. (about.fb.com) Another part of the sales pitch is reasoning, which is the industry term for having a model spend extra steps working through a problem instead of blurting out the first likely answer. Meta says Muse Spark can switch modes for harder questions and can send multiple software agents to work on parts of the same task in parallel, like several interns splitting up one research assignment. (about.fb.com) (techcrunch.com) The reason this launch is getting so much attention is that Meta spent the last year looking behind rivals. CNBC reported that Muse Spark is the first major model led by Alexandr Wang since he joined Meta nine months ago, after Meta made a $14.3 billion investment in Scale AI and put him in charge of Meta Superintelligence Labs. (cnbc.com) (reuters.com) That lab exists because Mark Zuckerberg was not happy with how Meta’s previous models landed. TechCrunch reported that Meta created Meta Superintelligence Labs after Llama models trailed OpenAI’s ChatGPT and Anthropic’s Claude, and Meta now says the new team rebuilt its artificial intelligence stack from the ground up over the last nine months. (techcrunch.com) (about.fb.com) Meta is not claiming a clean knockout over OpenAI, Google, or Anthropic. What it is claiming is a narrower gap with less compute, which is the expensive processing power needed to train and run these systems, and that tells you where the company thinks the next fight is: not only who has the smartest model, but who can afford to put one in front of billions of people. (cnbc.com) (axios.com) There is also a practical reason Meta wants a leaner model. A chatbot on a website can be slow and costly, but an assistant inside Instagram, WhatsApp, Messenger, and smart glasses has to respond fast enough that people keep using it, and Meta says Muse Spark is built to power exactly those products. (about.fb.com) The split inside Meta’s artificial intelligence strategy is now out in the open. The company is still associated with open-weight Llama releases, but its newest flagship effort is a proprietary model tied to Meta accounts, Meta apps, and a limited partner preview, which means the company now wants both the developer goodwill of open releases and the control of a closed product stack. (cnbc.com) (about.fb.com)