Meta has unveiled Muse Spark, its first AI model from the new Meta Superintelligence Labs unit, positioning it as a step toward “personal superintelligence” that can reason, use tools and orchestrate multiple agents on behalf of users. Announced on April 8 and highlighted by X account Coin Bureau as “Meta’s first step toward personal superintelligence,” the model is already live on meta.ai with a private API preview rolling out to select partners. Meta says Muse Spark is “small and fast by design, yet capable enough to reason through complex questions in science, math, and health,” and will gradually expand to WhatsApp, Instagram, Facebook, Messenger and Ray‑Ban AI glasses.

The launch follows nine months of rebuilding Meta’s AI stack under the Superintelligence Labs banner, created after Mark Zuckerberg vowed to “put superintelligence in the hands of everyone” and catch up to rivals such as OpenAI and Google. According to Meta’s blog, the company reworked its architecture, optimization pipeline and data curation to achieve similar capabilities to its previous Llama 4 Maverick model with “over an order of magnitude less compute,” describing Muse Spark as “the first step on our scaling ladder.” In comments reported by the Financial Times, Meta is explicitly leaning into niche strengths like health, arguing that medical and wellness queries are “one of the top reasons people turn to AI” and a space where differentiation from generic chatbots is possible.about.

Central to Muse Spark is a new “Contemplating” mode that runs multiple agents in parallel before responding, a feature Meta pitches as the answer to the “deep thinking” modes in frontier models like Gemini Deep Think and GPT Pro. AI at Meta described the mode as orchestrating several agents to reason together, achieving 58% on Humanity’s Last Exam and 38% on the Frontier Science Research benchmark, performance that the team says lets Spark “compete with the extreme reasoning modes” of its rivals. The model is also natively multimodal, able to process and generate both text and images, call external tools and manage sub‑agents to break down complex tasks, from financial modeling to troubleshooting household appliances.

Industry observers see the move as an attempt to reinsert Meta into the front rank of AI model providers after a period where most attention went to OpenAI, Google and Anthropic, similar to past catch‑up pushes that reshaped its core products. In a recent piece, Bloomberg noted that big tech firms are betting billions on proprietary “reasoning” models with agentic capabilities, framing Meta’s pivot away from purely open‑source Llama releases as part of a broader trend toward closed, vertically integrated AI stacks.

Beyond reasoning benchmarks, Meta is aggressively marketing Muse Spark’s health capabilities, calling it “one major application of personal superintelligence” and emphasizing that the model can generate interactive explanations of nutrition, exercise and medical information. Meta says it “collaborated with over 1,000 physicians to curate training data that enables more factual and comprehensive responses,” and external evaluations cited by the company claim Muse Spark scored 42.8% on the HealthBench Hard benchmark, outperforming rival models including Gemini 3.1 Pro and Opus 4.6, and slightly beating GPT‑5.4. Reuters reported that in practice Muse Spark can already estimate the calories in a meal from a photo and overlay objects, such as a mug on a shelf, demonstrating its combined visual and reasoning capabilities.

However, analysts quoted by the Financial Times and other outlets warn that turning a social network into a quasi‑medical assistant could trigger regulatory scrutiny, especially in the U.S. and EU, where health advice and data privacy are tightly regulated. At the same time, trading platform Invezz noted that Meta’s stock jumped around 9% on the day of the announcement as investors bet that a stronger AI strategy could drive new revenue streams and improve margins across Meta’s apps and hardware lines. For now, Muse Spark’s mix of agentic reasoning, health focus and multimodal capabilities suggests Meta plans to compete less on raw model size and more on targeted, high‑value use cases that can lock users into its ecosystem.

A group of North Korean IT workers made more than $3.5 million in just a few months by faking their identities to work as developers while also attempting to hack crypto projects, according to documents obtained by a hacker who compromised one of their devices.

The leaked data obtained by the unnamed hacker was shared by blockchain sleuth ZachXBT in a post to X on Wednesday. It revealed that one of the IT workers, “Jerry,” and a team of 140 members were making roughly $1 million a month, amounting to $3.5 million worth of crypto since late November.

The North Korean IT workers coordinated payments on a website called “luckyguys.site” using a shared password, “123456,” ZachXBT said, adding that some of the users on that platform appeared to work for Sobaeksu, Saenal and Songkwang, which are sanctioned by the US Office of Foreign Assets Control.

These crypto payments were converted into fiat and sent to Chinese bank accounts via online payment platforms like Payoneer. Tracing these wallet addresses also revealed links to other known North Korean wallets that were blacklisted by Tether in December, ZachXBT said.

Bad actors from North Korea and other countries continue to threaten the crypto industry with increasingly sophisticated tactics for carrying out hacks and scams. 

North Korean state-backed workers have stolen over $7 billion in funds since 2009, with a large share of that coming from crypto projects. The $1.4 billion hack of crypto exchange Bybit and the $625 million Ronin bridge hack are among its most notable attacks.

North Korean hackers were also blamed for the $280 million hack of the Drift Protocol on April 1. 

North Korean IT workers had a leaderboard

The North Korean IT workers who had their data exposed had a leaderboard showing how much crypto each IT worker had brought in for the organization since Dec. 8, with links to blockchain explorer pages showing transaction details.

Another screenshot shared by ZachXBT showed that Jerry used an Astrill virtual private network to access Gmail, where he submitted several applications for full-stack developer and software engineer roles on Indeed.

Related: Alleged Huione money-laundering boss extradited to China

In an unsent email, Jerry wrote a letter for a WordPress content and search engine optimization specialist position at a T-shirt company in Texas, seeking $30 an hour with availability of 15 to 20 hours a week.

Identification documents were falsified, too, with one of the IT workers, “Rascal,” sharing pictures of a billing statement using a fake name and fake address in Hong Kong. 

Rascal also shared a picture of an Irish passport, though it is not clear if it was used.

ZachXBT however said these IT workers were less sophisticated compared to other North Korean groups like AppleJeus and TraderTraitor, which “operate far more efficiently and present the greatest risks to the industry.”

Magazine: Asia Express: Phantom Bitcoin checks, China tracks tax on blockchain