Mark Zuckerberg's social media empire finally has something tangible to show for its artificial intelligence spending spree, though the distance between Meta and the frontier labs remains considerable.
Meta (NASDAQ:META) on Wednesday launched Muse Spark, the first major AI model produced since the company's US$14.3 billion investment in Scale AI brought former chief executive Alexandr Wang on board nine months ago.
Code-named Avocado, the model was built by Meta Superintelligence Labs - the unit Wang now oversees - and represents the most tangible product of the reorganisation Meta undertook after its Llama 4 family flopped with developers last April, prompting Zuckerberg to overhaul the company's entire AI strategy.
Shares rallied more than 9% in midday trading as investors digested the release alongside a broader market bounce tied to geopolitical developments in the Middle East.
Competitive, not dominant
Meta is not positioning Muse Spark as a frontier model, instead framing the release as a validation step for the architecture and training methods underpinning its new Muse series.
Published benchmarks show the model competitive with leading offerings from OpenAI, Anthropic, and Google across several task categories, without surpassing them across the board.
The company acknowledges persistent gaps in coding and long-horizon agentic systems - two areas where developer mindshare and enterprise contract value are most heavily concentrated right now.
Third-party evaluations from Artificial Analysis scored Muse Spark at 52, placing it behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6 on a composite ranking that tracks reasoning, multimodal perception, and tool use.
Llama 4's Maverick and Scout variants scored 18 and 13 respectively on the same index, so the step-up is material even if the social media giant still trails the leading pack by a margin.
The capex question
The four major hyperscalers are collectively projected to spend close to US$700 billion on capital expenditure in 2026, nearly doubling the historic levels reached in 2025.
Meta's guided range sits between $115 billion and $135 billion for the year, roughly twice what the company spent on capex in the prior twelve months.
Barclays analysts estimate that trajectory could drive a near-90% compression in the social media giant's free cash flow over the same period.
Morningstar analyst Malik Ahmed Khan noted that after a year of aggressive hiring and no major model releases, the company needed to demonstrate it was building something of substance for investors.
Muse Spark gives Wall Street a data point on execution, but the return-on-investment timeline for Meta's broader AI infrastructure buildout remains far from clear.
Distribution over disruption
Muse Spark will not be open-source - a deliberate break from the Llama playbook that signals Meta sees proprietary model access as a monetisation lever going forward.
The model now powers Meta AI's standalone app and website, with rollouts across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban smart glasses planned in the coming weeks.
A paid API for third-party developers is currently in private preview with select partners, though Meta has not disclosed pricing or a broader availability timeline.
Citizens analyst Andrew Boone noted Meta's structural advantage is its 3 billion-plus monthly active users, and that the commercial opportunity centres on advertising yield rather than competing for developer share against OpenAI and Anthropic.
Those two pure-play labs now carry a combined private valuation north of $1 trillion, and Anthropic's annualised revenue run rate hit $30 billion in April 2026, overtaking OpenAI at $25 billion - a sign of how rapidly the model-as-a-service market is consolidating around dedicated AI companies rather than platform conglomerates.
