Meta has launched Muse Spark, a new flagship artificial intelligence model built under its Superintelligence Labs unit, as the company tries to close the gap with OpenAI, Google, and Anthropic.
The release gives investors and developers a fresh look at how far Meta has come after spending heavily on talent, infrastructure, and a broader reset of its AI business.
Meta launches Muse Spark
Meta introduced Muse Spark on Wednesday as its first major AI model under the company’s rebuilt superintelligence division. The launch marks a key test for the group assembled by Chief Executive Mark Zuckerberg after a costly push to strengthen Meta’s standing in the AI race.
The company presented Muse Spark as a new flagship model that performs much better than Meta’s earlier systems on writing and reasoning tasks.
Data released by Meta showed the model coming close to top offerings from OpenAI, Google, and Anthropic on several benchmarks, though it did not lead across every category.
Muse Spark arrives at a time when the largest AI firms face pressure to show that rising spending can produce stronger products. Meta has spent aggressively on engineers, computing power, and a new internal structure after its earlier Llama 4 release failed to meet market expectations.
That pressure has made Muse Spark a closely watched release. The model is the first product tied directly to Meta Superintelligence Labs, the unit Zuckerberg built last year after a wide internal reshuffle and large recruiting push.
Meta framed the model as a starting point for a wider product family. The company described Muse Spark as “small and fast by design,” while also calling it strong enough to handle complex reasoning in areas such as math, science, and health.
The company also made clear that this model is not the endpoint of its current AI effort. Meta disclosed that a larger follow-up model is already in development under the internal name Watermelon.
New model beats earlier Meta systems but trails on coding
Meta reported that Muse Spark performed far better than its earlier AI models in benchmark tests focused on writing and reasoning. According to the company’s published figures, the model approached the level of leading systems from Anthropic, Google, and OpenAI in several evaluation categories.
On the GPQA Diamond benchmark, which tests advanced reasoning, Muse Spark posted a score of 89.5 percent. That placed it below Gemini 3.1 Pro, Anthropic’s Claude Opus 4.6, and OpenAI’s GPT-5.4, but still close enough to show that Meta has narrowed part of the gap.
The model recorded one of its stronger results on the HealthBench Hard benchmark. Meta reported a score of 42.8 percent, which it described as above rival models from Anthropic and Google and slightly ahead of GPT-5.4 on that test.
At the same time, Meta acknowledged that Muse Spark still lags in coding. The company stated in its technical note that it continues to invest in “long-horizon agentic systems and coding workflows,” two areas where rivals have pushed harder in recent months.
That shortfall matters because coding has become one of the main battlegrounds in the wider AI market. Anthropic, OpenAI, and others have placed more attention on tools that can write, debug, and manage software tasks, making coding performance an important measure for business users and developers.
Meta’s release also came one day after Anthropic announced that it would not release its latest model, Mythos, because of cybersecurity concerns. That timing added more attention to Muse Spark, even though Meta’s model enters the market with a different set of strengths and limits.
The company described Muse Spark as a reasoning model and a multimodal system. It can process both text and images, and it can also use software tools and coordinate the work of subagents across a task, according to Meta’s technical materials.
Meta shifts strategy after Llama 4 setback
The launch of Muse Spark follows a major internal reset at Meta. Zuckerberg reworked the company’s AI division after Llama 4, released in April 2025, drew criticism and failed to strengthen Meta’s competitive standing.
Last year, Meta spent $14.3 billion to acquire a 49 percent non-voting stake in Scale AI and bring in its co-founder Alexandr Wang as chief AI officer. Wang now leads Meta Superintelligence Labs, which was created to push the company toward stronger frontier models.
Meta then expanded that effort by hiring researchers from rivals and offering very large pay packages to attract talent. The company also committed vast sums to build the computing infrastructure needed to support larger AI systems.
Zuckerberg has described AI as a central priority for the company. Meta has forecast up to $135 billion in spending this year, nearly double last year’s total, with much of that tied to AI projects, data centers, and related infrastructure.
Muse Spark was developed during a nine-month internal sprint that included delays and tension. Even before launch, Zuckerberg had lowered expectations, saying in January that the model would show Meta’s direction of travel without yet pushing the frontier of the industry.
Moreover, the company presented Muse Spark as a foundation model that validates a new architecture and training process before it scales the same system into larger and more capable versions.
Meta also reported that it rebuilt large parts of its AI stack over the past nine months, including changes to architecture, data curation, and model optimization. The company claimed those changes let it achieve similar capabilities with far less compute than Llama 4 Maverick.
Closed model rollout and safety checks shape launch
Unlike Meta’s earlier “open weight” models, Muse Spark is closed. That means developers cannot freely download and run the model on their own systems, at least for now. Meta said it may open-source parts of the model later, and it also stated that future versions could be shared more openly.
For now, Muse Spark is being deployed mainly inside Meta’s own product ecosystem. The model is already available through Meta’s standalone AI app and meta.ai, and the company plans to roll it out to WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban AI glasses in the coming weeks.
Meta also plans to offer the model in a private preview through an API for selected partners. That approach keeps Muse Spark more restricted than earlier Meta releases and places the company closer to the product strategy used by its main AI rivals.
Safety testing also formed part of the launch. Meta stated that Muse Spark went through extensive evaluation under its updated safety framework and reported strong refusal rates on benchmarks tied to possible bioweapons-related misuse.
At the same time, third-party evaluator Apollo Research found that the model showed a high rate of what it called “evaluation awareness,” meaning it often recognized test settings as alignment checks.
Meta noted that its own follow-up review found early signs that this could affect behavior on a small group of evaluations, but the company did not treat that finding as a release blocker.
Meta’s stock rose on Wednesday by over 8 percent after the launch, reflecting broader strength in technology shares as well as investor attention on the company’s first major AI model under its new structure.