CALIFORNIA: Meta has launched Muse Spark, a new artificial intelligence model it says is the first step toward what it calls “personal superintelligence,” as chief executive Mark Zuckerberg makes his most ambitious push yet to close the gap with OpenAI and Google.
Muse Spark is the first model to emerge from Meta Superintelligence Labs, a division set up after Zuckerberg grew alarmed at the sluggish pace of the company’s AI development and the underwhelming performance of its Llama models against rivals such as ChatGPT and Claude. The model is available now on meta.ai and the Meta AI app, with a private API preview being rolled out to select users.
To signal how seriously it is taking the fight, Meta recruited Alexandr Wang, former co-founder and chief executive of Scale AI, to lead the new division, and invested $14.3bn in Scale AI for a 49 per cent stake. It has also aggressively poached researchers from OpenAI, Anthropic and Google.
Muse Spark is a natively multimodal reasoning model, meaning it can process and respond to both text and visual inputs. It performs particularly well on visual STEM tasks, entity recognition and interactive experiences, including creating simple games or helping users troubleshoot household appliances. Meta says it collaborated with over 1,000 physicians to sharpen the model’s health reasoning capabilities, enabling it to generate detailed, interactive displays on topics such as nutrition and exercise.
The company is also rolling out a feature called Contemplating mode, which deploys multiple AI agents working in parallel to tackle complex queries, allowing Muse Spark to compete with the most powerful reasoning modes offered by rivals such as Gemini Deep Think and GPT Pro. In benchmarks, it scored 58 per cent on Humanity’s Last Exam and 38 per cent on FrontierScience Research.
Not everything is straightforward. Users must log in via existing Meta accounts such as Facebook or Instagram, which is likely to revive long-standing concerns about data privacy. Meta has not explicitly confirmed whether personal data from these platforms feeds into the model’s training, though the company has a well-documented history of training AI on publicly available user data.
Meta says it conducted extensive safety evaluations before launch, including third-party assessments by Apollo Research, which flagged that Muse Spark showed the highest rate of “evaluation awareness” of any model it had tested, meaning the model frequently identified when it was being tested and adjusted its behaviour accordingly. Meta concluded this was not a blocking concern for release, but acknowledged it warranted further research.
Zuckerberg has promised progressively more powerful models ahead, including open-source offerings, and has made clear he wants AI that acts on behalf of users rather than merely answering their questions.
The race for AI supremacy is expensive, bruising and wide open. Meta has just told the world it intends to win it.