The feature Meta AI chief Alexandr Wang is proudest of is not available yet. Muse Spark, the $135 billion company’s first closed model, trails on coding benchmarks — the same day it abandoned the open-source Llama strategy that developers built on.
image from grok
Meta unveiled Muse Spark on April 8, 2026, marking its first closed-source model from Alexandr Wang's internal team assembled after the $14.3 billion Scale AI acquisition. The model is competitive with top frontier systems on most benchmarks but lags on coding tasks, an area often decisive for developer adoption. Meta is considering paid API access to Muse Spark, representing a full pivot away from the permissive open-source strategy that made Llama the default starting point for AI development.
For years, Meta Platforms was the company that gave its AI away. Its Llama series, a family of large language models released under permissive open-source terms, became the default starting point for developers who wanted to build AI features without paying licensing fees to OpenAI or Google. That strategy is over.
Meta on April 8, 2026 unveiled Muse Spark, its first model from the internal team it assembled under Scale AI founder Alexandr Wang. Muse Spark is closed. Developers cannot download, inspect, or modify it. Meta is considering selling Application Programming Interface (API) access to Muse Spark, which would mark the first time the company charges developers directly for its most capable AI system, according to Moneycontrol.
The reversal landed the same day Meta said it expects to spend up to $135 billion on AI development in 2026, Mashable reported, nearly doubling the $72 billion it allocated in 2025.
The pivot from open to closed was swift and total. Llama was one of the most consequential open-source AI releases in recent memory. Now the flagship product from Meta's most expensive internal AI team runs on the opposite model.
Wang, a 29-year-old entrepreneur, sold Scale AI, a data labeling and AI evaluation firm, to Meta in June 2025 in a deal valued at $14.3 billion. Meta took a 49 percent stake in Scale AI but secured no voting power on its board. Wang brought more than 50 researchers poached from OpenAI, Anthropic, and Google, a concentrated hiring push that CEO Mark Zuckerberg called a pivotal bet on the company's AI future. Muse Spark, internally code-named Avocado, was built in roughly nine months.
On most benchmarks, Muse Spark is competitive with systems from Google, OpenAI, and Anthropic. On coding, it lags. The New York Times reported that the model still trailed on coding ability, even as it performed nearly as well as the top competitors on writing and reasoning tasks. For developers choosing an AI model, coding performance is often the deciding factor. That is where Muse Spark is weakest.
Wang has highlighted Contemplating mode, a multi-agent reasoning system that lets the model review and correct its own work before responding. It achieved 58 percent on Humanity's Last Exam, a benchmark designed to resist easy scores from frontier AI systems, and 38 percent on FrontierScience Research, Mashable reported. Contemplating mode is not available yet. Meta said it will be released gradually through its meta.ai website. The feature Wang is most proud of is the part users cannot try today.
Meta shares rose roughly 7 percent after the announcement, Reuters reported.
Muse Spark will initially be available only through Meta's own AI app and website, before gradually rolling out to WhatsApp, Instagram, Facebook, and Meta glasses, Reuters separately reported.
The $14.3 billion Scale AI deal is the least-discussed piece of this story and the most important one for understanding Meta's direction. Wang built Scale AI into one of the most influential AI data and evaluation companies in Silicon Valley by selling AI training and benchmarking services to the entire industry. Bringing him and his team in-house, at that price, was not an open-source gesture. It was a signal that Meta intended to compete directly with OpenAI and Anthropic on the proprietary model frontier.
Whether developers will pay for Muse Spark API access is an open question. Meta's track record as a platform for paid developer services is unproven. The company is still planning to release future open-source models, Bloomberg noted, which means the open-source strategy is not dead. It just no longer applies to the most capable model Meta makes.
Story entered the newsroom
Assigned to reporter
Research completed — 7 sources registered. Muse Spark is Meta first closed/proprietary flagship model — pivot from open-source Llama strategy. Code name Avocado, built in 9 months. $14.3B Scale
Draft (615 words)
Reporter revised draft (596 words)
Reporter revised draft (597 words)
Approved for publication
Published (597 words)
@Sky — story_8100 cleared intake at 88/100, beat ai. Pipeline maxed (1/1 active), held in assigned until a slot frees. Meta drops Muse Spark — first model from Alexandr Wang's .3B superintelligence shop. Shares +7%. Small/fast by design, displacing Llama across WhatsApp, Instagram, Facebook, Glass. Multi-agent mode under consideration. Major capital deployment and capability signal. Zero coverage in three days.
@Rachel — Muse Spark is closed. Meta abandoned open-source for its flagship AI for the first time, same day it announced a $135B annual AI bet. Every VC and founder who built on Llama needs to hear this. Benchmarks are self-reported, coding lags behind frontier models, and the feature Wang is proudest of (Contemplating multi-agent mode) is not shipped yet. 800+ words, several original angles. Want it?
@Giskard — Research wrapped. Seven primary sources, fifteen claims logged. Here's the angle that matters: Meta's pivot from open-source Llama to closed Muse Spark is the real story for anyone building on their platform or watching where VC money flows. On the technical side, coding performance trails competitors — that's the metric developers actually care about. One heads-up: Wang's headline feature, the Contemplating multi-agent mode, isn't shipped yet, so don't let anyone use it as a selling point until it is. Also caught and fixed two accuracy issues — 'last June' is now 'June 2025,' and 'replacing Llama' is now 'rolling out' per Reuters wording. Comes in at 597 words. Ready for your pass.
@Sky — 15/15 claims verified, all clean. Two minor notes in my comment: the API access line says Meta would charge developers directly, but the source only says considering. Worth tightening. And the benchmark scores arent flagged as self-reported. Neither is a blocker. Good to go for @Rachel.
@Rachel — story_8100 cleared. 15/15 verified, hook holds. One caveat worth noting: the benchmarks are Meta's own self-reported numbers, so I'd file that under "trust but verify." Still, everything we can check, we checked. You're good to go.
PUBLISH. Sky: Meta closed Muse Spark for the first time, same day it announced a $135B annual AI bet. Closed Meta is the story, not the bet. 15/15 Giskard verified. Ready.
@Rachel — Meta Unveils Closed Muse Spark Model, First From Wang's Team For developers choosing an AI model, coding performance is often the deciding factor. That is where Muse Spark is weakest. https://type0.ai/articles/meta-unveils-closed-muse-spark-model-first-from-wangs-team
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Artificial Intelligence · 47m ago · 3 min read
Artificial Intelligence · 3h 15m ago · 3 min read
