Mark Zuckerberg once bet big on open-source models to propel Meta to the front of artificial intelligence. That wager faltered. Llama 4 landed with a thud. Developers shrugged. Rivals pulled ahead. So the chief executive shook up his team, spent billions more and handed the reins to a 28-year-old prodigy from Scale AI.
Alexandr Wang arrived in 2025 with a mandate to move at wartime speed. Less than a year later his group delivered Muse Spark. The model powers voice chats that interrupt and switch topics. It scans your camera feed in real time. It suggests Reels or maps while you talk. Yet questions linger. Can one flashy release erase the ground lost to OpenAI, Google and Anthropic?
The Ars Technica report paints a picture of frenetic activity inside a secure corner of Meta’s Menlo Park campus. Wang’s TBD Lab requires special badges. He and Zuckerberg keep offices there. The young leader recruited top talent at multimillion-dollar salaries. He fostered a startup vibe with boba tea happy hours. He argued that a small crew of exceptional engineers could outpace lumbering bureaucracies.
Some veterans felt sidelined. One researcher left for OpenAI after seven months. Others bristled when Wang described Muse Spark as built from scratch. The new model actually drew on Llama 4 code and datasets. Tensions flared. Still, the CNBC coverage from April 8, 2026 noted that Muse Spark showed competitive results on science, math and health reasoning tasks while using far less compute than its predecessor.
Meta poured money into the effort. Capital expenditures on AI could reach $135 billion in 2026, nearly double the prior year. Investors watched closely. They wanted proof the spending would lift revenue, not just burn cash. Zuckerberg signaled confidence. He told audiences that every ad on Meta platforms could be AI-generated by the end of the year. Give the system a product image, a budget and preferences. It writes copy, shoots video, picks audiences and optimizes delivery.
The shift from open to closed marked a stark reversal. Llama models once defined Meta’s generous approach. Chinese labs such as DeepSeek and Qwen then surpassed them on key benchmarks. Llama 4 failed to captivate. Behemoth, the much-hyped flagship, slipped. As Zapier explained in its recent analysis, Meta quietly retired the Llama herd in favor of proprietary systems. Muse Spark sits at seventh on independent leaderboards. It trails GPT variants, Claude Opus and Gemini but beats several strong contenders. Its visual skills shine. Users can point a camera and ask practical questions. Calorie counts from meal photos. Product recommendations from store shelves.
Yet gaps remain. Insiders say the model trails in long-horizon coding and agentic planning. Many Meta engineers still reach for Anthropic’s Claude when writing software. Wang pushed for deeper model work over rushed product features. In internal meetings he emphasized scaling laws and next-generation architectures. Product teams wanted faster rollouts across Instagram, WhatsApp and Ray-Ban glasses.
Meta Superintelligence Labs now steers the ship. Wang leads it. The group absorbed pieces of the old AI organization after restructurings and layoffs aimed at offsetting the massive spend. Employees once protested tracking software used to gather training data. Meta dialed back parts of the program after backlash. Trust inside the building has been tested.
Still, Muse Spark brought tangible upgrades. Voice conversations feel more natural. Users interrupt, change subjects or switch languages. The AI generates images on the fly and pulls context from social feeds. A new Incognito Chat mode on WhatsApp promises true privacy. Conversations vanish. No logs linger on servers. Inference runs in secure hardware enclaves. Zuckerberg himself touted the feature in May as essential for discussing sensitive topics.
Analysts remain divided. Some credit Wang with assembling one of the strongest research teams in short order. Russ Salakhutdinov, a former Meta AI vice president, called the output impressive. “Alex knows what he doesn’t know and he’s willing to listen,” he said. Meta’s official statement praised Wang’s record. “In less than a year, he’s helped build one of the strongest research teams in the industry and led Meta Superintelligence Labs as it launched Muse Spark.”
Critics see hype outrunning substance. One former employee told Ars Technica the bar had been set low both internally and externally. “The other labs are moving fast.” Incremental gains receive fanfare. Frontier breakthroughs stay elusive. Wang attended a White House dinner with President Donald Trump and fellow tech leaders. The moment underscored his rising influence. It also highlighted how personal relationships now shape AI strategy at the highest levels.
Meta experiments with new revenue paths. A limited API reached select partners. Paid access could follow. The company hopes third-party developers will build on Muse Spark and drive usage across its 3.5 billion monthly users. Shopping assistants, business agents, digital avatars and wearable experiences all sit on the roadmap. Success hinges on turning those experiments into habit-forming products before rivals widen their lead.
And the spending continues. Data centers. Chips. Talent. Nuclear power deals to feed the hunger for electricity. Meta’s stock rose after the Muse Spark debut. Markets gave the benefit of the doubt. But patience has limits. If the next models do not deliver clear superiority in coding, autonomous agents or video generation, pressure will mount. Zuckerberg has bet the company’s future on artificial intelligence. Wang holds the tactical command.
So far the results show promise without dominance. Muse Spark improves Meta’s own products. It narrows some gaps. It does not yet redefine the race. The coming months will test whether this expensive reset produces a true contender or simply keeps the social media giant in the conversation. Industry watchers will judge not by press releases but by developer adoption, user engagement and benchmark leaps that no one can dismiss.
WebProNews is an iEntry Publication