Meta has unveiled Muse Spark, the primary AI mannequin produced by its Meta Superintelligence Labs, the brand new AI analysis unit it created final yr and has spent billions of {dollars} to employees and equip.
The mannequin is, in line with benchmark checks that Meta printed, aggressive with main AI fashions from OpenAI, Anthropic, and Google throughout many duties, though it doesn’t surpass them throughout the board. Nonetheless, if the benchmark outcomes maintain up when examined by unbiased consultants, Muse Spark appears to place Meta again within the AI race after its final AI mannequin, Llama 4, which was launched in April 2025, was extensively panned as a dud.
Previously, nevertheless, Meta has been caught manipulating the printed benchmark outcomes of an AI mannequin to make it seem extra succesful than the model obtainable to most customers really was. This was the case with Meta’s Llama 4 benchmarks, during which the corporate later admitted to utilizing specialised, unreleased variations of the mannequin, fine-tuned for particular duties, to spice up benchmark scores in these areas, whereas the overall model made obtainable to all customers didn’t carry out as nicely.
And there’s one other catch. Few individuals will be capable to use the brand new Meta mannequin exterior of the corporate’s personal product ecosystem. Not like Meta’s earlier AI fashions, which have been launched as “open weight” fashions—that means anybody might obtain the fashions at no cost and run them on their very own gear, in addition to modify and fine-tune them as they wished—Muse Spark is, no less than for the second, primarily an in-house device for Meta.
The mannequin at the moment powers the Meta AI assistant within the firm’s stand-alone Meta AI app and on meta.ai. The corporate mentioned it will likely be rolling it out to WhatsApp, Instagram, Fb, Messenger, and Meta’s Ray-Ban AI glasses within the coming weeks. It additionally mentioned it would provide the mannequin in a “private preview” to pick out companions by way of an software programming interface (API). That makes Muse Spark much more proprietary than the paid proprietary fashions supplied by Meta’s rivals. (Meta mentioned in a weblog submit that it hopes to open-source future variations of the mannequin.)
Muse Spark is Meta’s first reasoning mannequin, that means it could work by way of a course of in a step-by-step trend, utilizing completely different methods if its preliminary strategy doesn’t work. The corporate’s earlier fashions have been all designed to provide an instantaneous reply based mostly on the mannequin’s coaching. Muse Spark can also be a multimodal mannequin that may absorb and output each textual content and pictures. The mannequin additionally helps the usage of different software program instruments and can assist orchestrate the work of a number of subagents, in line with a technical weblog submit launched by Meta.
In its weblog submit asserting the brand new mannequin, Meta describes Muse Spark as “small and fast by design, yet capable enough to reason through complex questions in science, math, and health.” It describes the mannequin as the primary in a sequence of latest fashions, with Muse Spark getting used to validate the structure and coaching regime Meta is utilizing, earlier than the corporate scales this as much as bigger and much more highly effective fashions in the identical household.
The mannequin additionally has a “contemplating” or “thinking” mode during which it could spin up subagents to motive about completely different elements of a job in parallel. Meta mentioned in a technical weblog it printed on the brand new mannequin that this mode permits Muse Spark “to compete with the extreme reasoning modes of frontier models such as Gemini Deep Think and GPT Pro.”
The benchmark outcomes printed alongside the launch paint an image of a mannequin that’s aggressive however not dominant. For example, on the GPQA Diamond benchmark, which is meant to check PhD-level reasoning talent, Muse Spark scored 89.5%, which barely trailed each Gemini 3.1 Professional’s 94.3% in addition to the 92.7% and 92.8% that Anthropic’s Claude Opus 4.6 and OpenAI’s GPT-5.4 scored respectively. On a number one well being benchmark, HealthBench Arduous, Muse Spark beat all rival fashions with a rating of 42.8%, which was much better than both Opus 4.6 or Gemini 3.1 Professional, and barely higher than GPT-5.4.
Meta acknowledged the efficiency gaps. Its technical weblog submit states that the corporate continues “to invest in areas with current performance gaps, specifically long-horizon agentic systems and coding workflows.”
The Muse Spark launch is essentially the most tangible product but of the sweeping reorganization Meta undertook after the Llama 4 fiasco. In June 2025, Meta spent $14.3 billion to amass a 49% nonvoting stake in Scale AI and introduced in its cofounder and CEO, Alexandr Wang, as Meta’s first-ever chief AI officer.
Wang has been tasked with main a newly created Meta Superintelligence Labs unit. Wang and Zuckerberg went on a expertise acquisition spree, providing AI researchers at rival AI labs pay packages that reportedly climbed into the a whole lot of hundreds of thousands of {dollars} when fairness was included. The corporate has additionally dedicated a whole lot of billions of {dollars} to construct out AI computing infrastructure to help its new AI drive.
There has since been additional reorganization, whilst Muse Spark was in improvement. In March 2026, Meta created a brand new utilized AI engineering group led by Maher Saba, a vice chairman who beforehand labored in Meta’s Actuality Labs digital and augmented actuality unit. Saba experiences on to Meta chief expertise officer Andrew Bosworth. Saba’s unit works alongside Wang’s Superintelligence Labs to construct what an inside memo described as “the data engine that helps our models get better, faster.” The transfer was extensively interpreted as Zuckerberg hedging his bets—guaranteeing product-focused AI improvement continues whilst Wang pursues longer-term superintelligence analysis.
In a technical weblog submit, Meta says that over the previous 9 months its workforce rebuilt its AI stack from the bottom up, together with enhancements to mannequin structure, optimization, and information curation. The corporate claims these advances enable it to attain the identical capabilities with “over an order of magnitude less compute” than Llama 4 Maverick, Meta’s earlier mannequin. Meta additionally says its reinforcement studying pipeline now delivers “smooth, predictable gains,” and that Muse Spark is step one on a deliberate “scaling ladder” the place every technology validates the final earlier than the corporate trains bigger fashions.
On security, Meta says Muse Spark underwent in depth analysis earlier than deployment, following the corporate’s up to date security framework. The mannequin experiences spectacular outcomes for security round potential bioweapons engineering—on one benchmark, it refused 98% of requests that the benchmark designers judged as doubtlessly serving to somebody develop a bioweapon.
Nonetheless, the weblog submit additionally mentioned third-party evaluator Apollo Analysis discovered that Muse Spark demonstrated the very best price of “evaluation awareness” of any mannequin Apollo has noticed, incessantly figuring out check eventualities as “alignment traps.” Meta says its personal follow-up investigation discovered preliminary proof that this consciousness might have an effect on mannequin habits on a small subset of alignment evaluations, however concluded it was “not a blocking concern for release.”
In 2001, Fortune first convened “The Smartest People We Know,” bringing collectively CEOs and founders, builders and traders, thinkers and doers. Since then, Fortune Brainstorm Tech has been the place the place daring concepts collide. From June 8–10, we’ll return to Aspen—the place all of it started—to mark 25 years of Brainstorm. Register now.
